Amazon Finds Solutions At The Crossroads Of AI And Economics

Any sound economic policy in any country would need appropriate measuring techniques to keep track of its inflation. This is because these inflation indices can further be used for aggregating productivity, cost of living and other attributes that help in decision making to avoid any catastrophe. The same can be said of online consumer businesses. E-commerce giants like Amazon have to keep track of all the trends to have a relevant stock and cost options. However, they are a few challenges when it comes to online consumer business:

The global trade environment (millions of products)
Frequent price changes
Extremely high turnover for some products (e.g., apparel, electronics)

To address these challenges, researchers at Amazon has started to develop methods that utilize AI and econometrics tools to predict quality-adjusted prices using text and image embeddings.

Here’s how ML models can be used to keep a track of the inflation:

Train the model on data from one year.
Feed descriptions of products on the shelves a year later.
The model will predict the prices of products according to the earlier valuation.
Comparing the predicted and actual prices provides a measure of inflation.

“It’s easy to get siloed doing engineering, machine learning, natural language processing, computer vision, stats, operational research, and, economics. But, when these disciplines interact, you get more interesting and useful results.”

Pat Bajari, Chief Economist @ Amazon Core AI

Need for Hedonic Pricing In The Age Of Deep Learning

The housing market is the most common example of the hedonic pricing method, wherein the price of a building or piece of land is determined by the size or locality and not the housing prices of the region. In the case of online consumer businesses, this hedonic pricing is needed to accomplish the following:

To avoid biases in the matching set, we can predict the prices of missing products during the period-to-period comparisons to find relevant product categories with high turn-over.
In product groups like apparel, about 50% of products get replaced with new products every month
Matching sets can be non-representative of good baskets, creating systematic biases
Use predicted prices and product attributes or qualities, instead of the missing and observed prices.

To predict a product’s price, the team at Amazon fed their machine learning model with data such as the number of reviews, average star rating, and other textual data such as product descriptions and titles, along with the visual data such as product shots.

To validate the significance of this model, conventional statistical approaches and deep learning networks are compared. And, the results show that neural networks and linear regression have underpredicted the price. Neural networks predict the price of $230 whereas linear regression predicts the price of $90 when the actual price is $2000. The designer dress used for validating models is a hard item to predict because few attributes can be very intricate —for example, custom-made rose gold buttons.

The basic mathematical framework for time series forecasting is a century old, but the scale of modern forecasting problems calls out for new analytic techniques.

The above-discussed models can be also used to assess business trends. But, the economists at Amazon Core AI team believe that if the same approaches are used by the bankers to represent products of the economy as a whole, they could observe inflation rate variations in real-time.

“If central bankers have a view with a one-day latency, it could give them signals about whether monetary policy is too loose or too tight,” Bajari explains.

Not only pricing but the models can also be used in monitoring the inventory through randomised experiments.

Avoiding An Online Spillover With AI

Randomized experiments are popular in drug testing and many other industries in order to assess how effective a strategy or solution is. In case of a drug test, two groups are considered — where one group is given the real medicine while the other one has been given something else as a placebo effect. If the group which had the medicine show significant improvement then the drug can be deemed effective.

Now, in a similar fashion, taking an instance where a certain treatment has resulted in faster delivery of certain products to a group, which in turn has resulted in the rising popularity of those products. In such a case, the recommendation engines at Amazon will start to recommend these popular products more frequently, even to the customers of the other group. This is called a spillover effect where the actions of one group are affecting the other group, unlike in the case of drug testing.

This happens because of the complex feedback loops set in place by the consumer industries. Pat Bajari and his peers, in their work, have recommended multiple randomizations to identify causal relationships underlying the data.

To assess the trade-offs of deploying a machine learning model in order to curb spillover effects, inventory management, the Amazon core AI team believes that the expertise of an economist is crucial.