My eye was drawn to an article in the New York Time last week – For $2 a Star, an Online Retailer Gets 5-Star Product Reviews. This article drew attention to the ongoing and growing problem of fake reviews. Like many of us I increasingly rely on reviews on sites like amazon.com or yelp.com to guide purchasing decisions, especially for products/services/towns about which I know nothing. Clearly fake reviews, reviews essentially paid for, are an issue. How then would a Decision Management System help you address this? At one level it is a classic fraud problem – how can you flag those transactions that are fraudulent so that automated systems don’t simply allow them and so limited human resources can be applied to investigate those most likely to be problematic? 6 Steps seem called for:
- Discover the relevant Decisions
The first step is to determine the key decisions. Clearly we are deciding “Should this review be accepted?” and this decision is dependent on several lower-level decisions such as “Can we trust this reviewer?” to check for problem reviewers, “Can we trust this company/product?” to check for companies who are repeat offenders and “Does this review seem genuine?” Only with all these dependent decisions answered can we realistically decide if we should reject the review. This decomposition is at the core of the decision discovery process.
- Automate these decisions
Clearly for most online sites any attempt to catch this fraud has to be automated. This means automating the decision so that it can be applied 24×7 to the high volume of reviews a typical site receives. Creating a simple interface that can be called before a review is posted and hooking that up ready for development is next.
- Capture expert and policy rules
The easiest way to define how to make these decisions is to capture specific, expert rules for each decision. These rules might come from hard-won experience, observing problems that other companies have, getting advice from the Feds or other outside groups or be a matter of policy. I suppose there could also be some from regulations in the end. These rules would need to be able to access the properties of the review (reviewer, data, IP address) as well as its content.
- Data mining for additional rules/better thresholds/outliers
One of the best ways to build on these kinds of rules would be to use data mining techniques – either “real” ones in a data mining or predictive analytic workbench or “soft” data mining using visualization and reporting tools to allow an experienced business user to seek patterns in the data. Given the nature of reviews this would likely involve text mining, looking for sentiment or phrasing that seems fake or repetitious for example. This analysis might result in new rules that can be applied or might change the thresholds or conditions of a rule to be more statistically significant. It might also create rules you would never have thought of and help you identify the kinds of outlier behavior that is so indicative of fraud.
- Predictive analytics
The next step would be apply true predictive analytic models that take the characteristics of the review and predict the likelihood that it is fraudulent (like the model the professor in the article is seeking). Again text analytics would be key but one should not underestimate the power of structured data (number of reviews, sudden changes in the kind of reviews being recorded and so on). These scores would likely be combined with other features in new rules such as “if the review pattern fraud score is above this and the reviewer has previously been flagged for a suspicious review then….”
- Next best case for review
By this point we are managing a pretty sophisticated decision with expert rules, data mining and predictive analytics all playing a part in making good decisions. At this point we probably want to think about a new decision – “What is the best case for an available staff member to investigate?” This Next Best Case mindset replaces a FIFO case management environment with one that applies the risk of fraud, the potential impact of the fraud and perhaps other criteria to select the cases to investigate dynamically as staff members become available. Lower volumes are involved but the impact of improving this decision can be significant.
- Decision Analysis
Last but by no means least we must institute the kind of ongoing decision analysis that will allow us to stay ahead of fraudsters. Constantly identifying new patterns, spotting the ways in which companies are gaming our existing checks and generally engaging in an arms race with those who would commit fraud is essential. You are not done with this or any other Decision Management System when it is first built. Ongoing monitoring and continuous improvement are key.
So there you have it – my prescription for dealing with this (or really any other) kind of fraud.
By the way, these stages are all described in more detail in Chapters 5-7 of Decision Management Systems. If you are dealing with fraud and would like to walk through how this approach would work for you, drop me a line firstname.lastname@example.org and I would be happy to chat about it.