The advanced technology used by companies like Amazon and Netflix can be utilized by tax administrators to make better decisions faster. A recommender system is a tool that takes large and complex information and categorizes and prioritizes it for an individual user. A company like Amazon takes your demographic information and web traffic input to create a list of associated products you may also like to purchase. LinkedIn suggests connection that are relevant to you instead of suggesting random people from its pool of 500 million users. Netflix knows you love action movies and is less likely to suggest a subtitled dramatic foreign film. These recommender systems increase basket size, purchase amount, user activity, and customer satisfaction.
A recommender system, at heart, is a simple concept, but to make it work for any industry,these systems incorporate several techniques such as:
- artificial intelligence
- machine learning
- data mining
- user modeling
- case-based reasoning
- constraint satisfaction
These systems are about increasing efficiencies in all aspects of business – and that includes tax administration. By combining human intuition and knowledge with AI driven recommender systems,tax administrators can arrive at solutions faster (think better audits, faster collections, and targeted enforcement). Decision making, comparisons, and exploration are found in every aspect of tax administration and can be enhanced by recommender systems. Recommender systems would allow for improvements in the provision of tax administration, as well as the taxpayer experience as the “customer”.
Traditionally, tax compliance has been driven by business rules. If-Then conditions that are woefully inadequate in identifying non-compliance. Instead of an auditor sorting through thousands of cases with little or no analytic backing for selection, recommender systems can segment the population into issue-focused groups of taxpayers (or returns) that are most likely to have one or more characteristics of non-compliance. The human element of this process is introduced when scores derived from the recommender system (sometimes a machine learning algorithm)are not overwhelmingly positive or negative. An examination takes place and a human decision is made on a specific return. How could we apply a recommender system to this scenario? We have a few key elements:
- User input:the decision auditors have given this specific return
- User information: the auditor notes associated with the case which can be text mined for additional information
- Product information: the return itself
- Demographic information: all the information the agency has on the taxpayer, and the other taxpayers and entities linked to them
Using these four elements, a recommender system could surface more returns that have similar risk assessments with similar characteristics and recommend a course of action that will quickly drive to the most likely outcome.Now, instead of the auditor hunting for similar taxpayers, the recommender system presents returns that have similar non-compliance tags. This eliminates the ‘search and find’ aspect of the job, increasing efficiency, and the new inputs from these recommended returns can help generate even better recommendations.