Dr. Rajhans Mishra
Information Systems Area, IIM Indore
Email: firstname.lastname@example.org; Phone: 0731-2439550
With the advent of web 2.0, a massive amount of data is being generated by the web users. Web users are not only consuming data but also become a great source of data generation. Users are generating data on various platforms explicitly and implicitly. For example, on social media platforms like Facebook and Twitter, users explicitly generate the data through their posts, comments, clicks and likes and implicit data generation happens during their navigation on different web pages, time spend on various activities as gaming or using a particular application. E-commerce platforms as Amazon and Flipkart also accumulate explicit data when customers are writing the reviews or rating the products, while implicit data is captured during customers interactions like browsing product pages, adding of products to the wish list and reading the reviews. Most of the platforms that captures the users’ data (directly or indirectly) try to capture their preferences in order to capture more value from the prospective customers.
Recommendation systems are most popular systems to help these platforms to capture the behaviour of the customers and suggest the offerings accordingly. The business model using recommendation systems may vary but the core objective to use recommendation systems remains the same, that is capturing the preference and suggesting accordingly. An ideal recommendation system will be able to sense precisely what is the next and immediate requirement of the user and accordingly will suggest the desired product or services. New age companies like Amazon and Netflix are massively using recommendation engines to generate additional revenue through up-sell and cross-sell strategies.
While browsing through web, users go through various web pages as per their requirements and preferences. Predicting the next page visit of the user at any point of time is a great insight with respect to the temporal movement of any user. It is equivalent to knowing the micro movement of the people in physical world that can be utilized for various business purposes. Similarly, knowing the sequential visits of the users will help for target marketing by focusing the relevant advertisements on the pages user is about to visit. It may be help in real-time pricing of advertisements on web pages and can create more value in the entire ecosystem of digital marketing. In case of e-commerce companies it may help to suggest a page (essentially containing information about a product or service) that is required at that point of time by the user.
In our work, we have designed and tested a recommendation system that can suggest the next page visit of any particular user after analysing his previous visits. The recommendation is based on its prediction on user preferences on web page visits. Web data exists in various formats like content available on website, hyperlink structure among websites, and usage / navigation pattern of the users. The proposed recommendation systems works on navigational data of the users that has been stored in web servers of the websites and search engines, accordingly it can be used for intra- website and inter-website recommendations.
On the basis of development methodology and philosophy, recommendation systems are primarily classified into content-based and collaborative systems. While content-based systems uses specific user’s information to generate the recommendation, collaborative systems uses the information of similar group of users for the recommendation. The proposed recommendation system in our work is a collaborative system that uses the collective wisdom of other users to generate the recommendation.
Our recommendation systems has four steps to generate the final recommendation for the users. The steps are as follows: (a) Clustering of user sessions (grouping of similar users on the basis of their navigation pattern). Since users may have multiple interests, hence overlapping clusters are allowed to capture the preferences through rough sets (b) Generating the response matrix from the user groups (c) Assigning the weights to the existing web navigated pages on the basis of sequential visits (d) Deriving the weights for non-visited pages using a mathematical technique, singular value decomposition.
Designed system is further tested using three datasets, out of which two datasets were real navigational patterns of users from specific websites and one was a simulated dataset. The benchmarking of the proposed recommendation system has been done with a random prediction model and existing similar kind of system. A rigorous testing has been performed for multiple subsets of the datasets and it has been found that performance of the proposed system was better that the random prediction model as well as existing sequence- based system. Hence, the viability of our approach has been justified by the experimental results. The proposed system can further be augmented by adding the contextual information which may further improve the efficiency of the prediction as well as will solve the cold start problem (prediction without having the initial few visits of the user). It can be well utilized as a product in real scenarios by platforms that are having data about users’ navigational visits to capture their interest.
ORIGINAL ARTICLE: Mishra, R., Kumar, P., & Bhasker, B. (2015). A web recommendation system considering sequential information. Decision Support Systems, 75, 1-10.