Minimizing Recommendation Fraud

Exploring the applicability of SingularityNET’s weighted liquid rank reputation algorithm in reducing recommendation fraud via experiments on simulated Amazon-like marketplaces.

Introduction

From the rising trend of deepfakes to the age-old quest of gaming the recommendation engines of major e-commerce platforms, the digital world has always been home to scams, fraud, and false information.

With the rise of online retail, and the resultant increase in the competition among suppliers, the financial incentives to game the system have only increased. Fraud is especially prevalent in online product recommendations, primarily because 82% of Americans check for product reviews before making a purchase, and that the search engines give prominence to products with better reviews.

The frequency and sophistication of the scams regarding online product recommendations are on the rise. Even Amazon, probably the most sophisticated internet retailer, has been unable to prevent a wide variety of fake product recommendation campaigns. Most of these scam campaigns function by incentivizing dishonest buyers to leave fake reviews and ratings, which help the inferior products gain more visibility through recommendation engines.

Recommendation scams pose a fundamental challenge to online commerce. One of the primary advantages of e-commerce is the ability to offer a wide variety of products and brands. Only a crowdsourcing approach can deliver the depth of product information and ratings that are required to navigate this mess of abundance. However, consumer reliance on crowdsourced online product reviews and ratings depends upon their perception that the crowdsourced information is accurate. Recommendation scams damage that perception.

Weighted Liquid Rank Reputation Algorithm

We are excited to announce that SingularityNET researchers have developed a suite of techniques to solve the recommendation fraud problem by making it more expensive and challenging to engage in such dishonest practices.

These techniques have been under development in SingularityNET’s research lab since early 2018 as part of our work on creating an integrated reputation system for SingularityNET’s decentralized AI marketplace. Such a reputation system will ensure that SingularityNET’s marketplace is fair and honest after it achieves a large scale.

The SingularityNET reputation system is based on an approach called the “weighted liquid rank reputation algorithm,” developed by SingularityNET’s AI researcher Anton Kolonin and other team members.

The same methods that are designed to work for SingularityNET’s large-scale AI marketplace are also applicable to other online marketplaces, even retail marketplaces like Amazon and its competitors. Recently, SingularityNET published an in-depth study on the behavior of digital marketplaces participants. The research was conducted to develop a reputation algorithm that can perform more effective product recommendations.

To explore the applicability of our work on the SingularityNET reputation system to the retail e-commerce case, we created several simulations of Amazon-like marketplaces, using real-world information cultivated from in-depth market research.

In our simulations, we were able to reproduce the quantity of manipulated reviews and financial rewards for scamming suppliers that exist on Amazon.com, and then evaluate how our SingularityNET reputation-management methods would be able to handle this sort of situation.

For the simulations described here, we configured a version of the weighted liquid rank algorithm to provide accurate recommendations for products listed at online marketplaces and then added new features aimed at easing scam prevention. While still preliminary in some ways, the results obtained are quite exciting and give some clear directions regarding how online marketplaces should manage recommendations and ratings to remain relatively impervious to scammers.

Simulations of Recommendation Fraud

The goal of SingularityNET’s reputation system is to minimize the amount of fraud that takes place and encourage thorough and accurate ratings. Our focus, therefore, is on configuring the overall system to promote positive and useful behavior, rather than on catching bad actors and squashing scams.

Since fake raters generally work hard to conceal their behavior, it is easier to estimate how factors like ratings distributions and review patterns affect the total amount of scamming, than it is to identify individual fake reviewers. One can minimize the total negative effects of scamming behavior systematically by reducing incentives for scamming and maximizing the costs incurred by scammers.

Our reputation system accounts for numerous factors, such as:

• the flow of reputation through time (so that reputation scores from earlier times influence future scores but in a judicious way);
• the flow of reputation from raters to ratees and also vice versa in the opposite direction;
• the amount of time various parties have been carrying out transactions in the marketplace;
• the correlation between payment flows and ratings.

The algorithm can be configured to suit different marketplaces with diverse market conditions and dynamics. Due to the very liquid and flowing manner in which our reputation system assigns and transfers reputation, we call it the weighted liquid rank reputation algorithm.

To explore the applicability of our weighted liquid rank approach, we created simulations of Amazon-type marketplaces. The simulations had up to 1000 simulated buyers, 10,000 simulated sellers, 1000 simulated products being bought and sold, and between 100–1000 days of marketplace activity.

In this setting, we simulated scamming vendors who sell low-quality products and who allocate some of their funds for purchasing ratings from the “gaming-ready” fraction of buyers, until they succeed in creating a market for their inferior products.

In our simulations, buyers purchase goods and services online by following their built-in utility parameters, with the purchase priorities increasing with the length of time between purchases. Sellers are assigned to supply these goods in proportion to demand to meet needs. Consumers remember scamming suppliers from which they made their purchases and avoid them. The consumers also remember the good suppliers, tending to return to them for their purchases.

For the sake of simplicity, in all of the simulations reported here, we have set the number of product categories that consumers need to five. One of these five product categories is utilized by dishonest suppliers to sell their inferior products. Lastly, we configured each supplier to sell an average of one hundred different products — all of them with the same price and product category.

In evaluating our simulated marketplaces, we use reputation scores, and profit amounts to determine how many scam campaigns dishonest buyers will purchase for the products they sell. We also looked at whether dishonest buyers will change a product that is not doing well, and whether dishonest buyers will change their identity.

Before running these simulations, we improved the basic version of the algorithm by adding additional parameters which can then be adjusted to suppress scamming effects optimally. We did this because the liquid rank reputation algorithm, in its simplest form, is not especially impervious to scamming and cheating.

We also developed multiple metrics for analyzing algorithm performance, such as:

● OMU: Organic Market Utility — This is the fraction of buys that legitimate buyers made from legitimate sellers based on ratings, as a proportion of the maximum amount of buys possible in the absence of ratings. We wish to maximize this metric.

● LTS: Loss To Scam — This is defined as the fraction of all spendings made by honest buyers paid to dishonest sellers. This metric shows what proportion of money spent by honest buyers is spent on the products offered by dishonest sellers, so we want to minimize it.

● BSL: Buyer Satisfaction Loss — This is calculated as the average of expected dissatisfaction on behalf of honest buyers weighted by financial values. We want to minimize BSL, so that honest buyers get the highest quality products possible for the price.

● SGP: Seller Gaming Profit — This is the total scam income (organic buys of products times dissatisfaction with each product) divided by the cost of scam (sum of all sponsored buys by dishonest buyers together with their commission rate). Low SGP values correspond to small profits by scamming sellers from honest buyers, so we wish to minimize it.

Finally, we added three new extensions to the previously defined algorithm:

● The “Anti-biased” feature takes into account buyers’ rating histories to solve the rater bias problem in which some raters rate items higher than do other buyers — either for natural reasons or if they are being paid for those ratings;

● The “Predictive” configuration of the reputation system relies on buyers’ review and rating histories to infer how accurate their ratings are in predicting reputations of sellers. In determining product ratings, we then weight buyers’ impacts, based upon the accuracy of those predictive capabilities;

● The “Vendor Impact” feature infers the reputations of the sellers providing multiple products from the reputation of each of the individual products and then uses the seller reputation to infer the reputations of other products offered by the same vendor. This is helpful in cases where some products have more reliable rating histories than others.

Simulation Results

The chart below displays the simulation performance in relation to a metric called “Probability of Leaving a Rating” (PLR).

Figure 1. Table and chart for the dependency of financial metrics on probability for leaving ratings for “Regular” reputation system using our simulations. The charts show 95% confidence intervals.

Improved metric values are positively correlated to higher PLR values. Since more honest buyers leave reviews and the same number of scam buyers leave reviews, we get more informed decisions.

It should also be noted that the “Regular” reputation system is considerably worse than no reputation system at all.

The performance of the financial metrics is shown below in the tabular data and charts of Figure 2.

We made simulations involving a variety of systems: Regular, Weighted, TOM (Time On Market), SOM (Spendings On Market), Anti-biased, Predictive, and Vendor Impact.

Also, we targeted optimization at maximizing OMU (Organic Market Utility) and minimizing the other metrics.

Figure 2. Table and chart are presenting the performance of financial metrics for different reputation systems using simulation. The charts show a 95% confidence interval for the highest and lowest the true values could be (had we repeated the simulations indefinitely).

Our simulation results show that in the “Anti-biased” configuration of the reputation system buyers lost less money to low-quality products, with gamed reputations, and were less dissatisfied with their purchases, as measured by LTS (Loss To Scam) and BSL (Buyer Satisfaction Loss).

One can also see that the “Anti-biased” configuration of the reputation system makes the scam the least profitable based on SGP (Scam Gaming Profit) column data.

Finally, the ability to distinguish honest sellers providing high-quality goods from dishonest reputation gamers is presented in the following charts, which display reputation ranks of all sellers on the market.

Figure 3. Distribution of reputation ranks for dishonest (red, left) and honest (blue, right) sellers when “Regular” reputation system is in use.

Figure 4. Distribution of reputation ranks for dishonest (red, left) and honest (blue, right) sellers when “Anti-biased” reputation system is in use.

The first chart refers to what we call the “Regular” configuration of a reputation system, and the second chart displays the “Anti-Biased” configuration.

The scamming sellers are on the left (red dots), and honest sellers are on the right (blue dots). The seller reputation ranks are computed as averages across reputation ranks of all products.

Conclusion

Our simulations let us conclude the following:

● The probability of leaving a rating greatly impacts the performance of our reputation system. When more real buyers provide feedback on their experiences, we can better determine product quality, which makes it easier to separate dishonest buyers from honest ones. From our simulation results, we strongly recommend that online marketplaces encourage buyers to leave ratings as often as possible.

● Using the “Anti-biased” reputation system configuration, with parameter settings as detailed in our business analysis report, provides a much stronger scam resistance in marketplaces with ratings design and behavior similar to those of Amazon.

These lessons may be practically useful right away to anyone running an Amazon-type marketplace.

As we further develop SingularityNET’s decentralized AI-services marketplace, we will follow a similar methodology — using simulation models (carefully tuned based on real-world market data) to experiment with critical reputation-system parameter values and drawing scientific conclusions regarding what combinations of costs and incentives yield an overall healthy marketplace with a minimum of scamming and other pathologies.

Markets such as these are complex self-organizing systems with rich dynamical complexities, but can still be understood and probabilistically controlled using appropriate scientific methodologies.

What’s Next?

We want to refer those readers who are interested in additional details to our business analysis report.

You can also meet with the SingularityNET team members during the poster session of the AI for Social Good Workshop at the IJCAI-2019 conference in August 2019.

If you would like to learn more about SingularityNET, we have a passionate and talented community which you can connect with by visiting our Community Forum. Feel free to say hello and introduce yourself here. We are proud of our developers and researchers that are actively publishing their research for the benefit of the community; you can read the research here.

For any additional information, please refer to our roadmap. Do not forget to subscribe to our newsletter to stay informed about all of our developments.