Blog: ML & AI: Building a More Responsive Supply Chain

By Dr. Chandler Johnson

GAINS Labs seeks to democratize data science by applying Machine Learning and Artificial Intelligence to supply chain challenges, addressing customers’ real business problems. Our data scientists work alongside customers to identify needs, outline research programs, and master supply chain planning.

Here are some examples of GAINS Labs programs I recently shared at the GAINS Summit illustrating the potential value of ML and AI-based solutions to supply chain planning challenges:

Learning Rules for Force-Multiplying Human Time

Humble spam filters are mundane, taken for granted, and generally ignored: until we see the thousands of emails managed for us when an intelligent system observes our actions and automatically learns decision rules. The mundane spam filter dynamically learns what matters and what doesn’t, and its efficient sorting becomes a force-multiplier for human time.

GAINS Labs’ data scientists are building systems that use similar tools to learn business rules. For example, can a learner independently propose experiments to a human, observe human responses, and then learn rules for curating problematic demand transactions? Can the same learner regularly retrain itself by proposing new experiments and observing new responses in regions where it is uncertain?

In one instance, we had 18 million demand transactions, of which 1.7 million were manually or heuristically excluded. The underlying heuristics had accumulated over ten years, and some were likely out-of-date. Because of the volume of transactions, newer, as-yet-unrecognized heuristics are certainly lurking in existing data, waiting to be discovered once the associated transactions start manifesting in problematic output.

GAINS Labs applied our experimental rule learner to this 18 million-row dataset, asking a machine to find 200 rows that the system understood to be “unusual.” We then asked a domain expert to review those 200 computer-selected rows and classify them as either problematic or not: Spam, yes, or no. With those labeled data, we trained a supervised routine to predict “unusual” rows for the remainder of the dataset. For 100 more computer-selected rows, we asked the domain expert to classify whether each row was problematic. We repeated that process once more and applied the learned rules to the entire dataset.

With only 400 human-labeled rows, the learner found more than 1.5 million of the 1.7 million “bad” rows. The learner misclassified 25,000 “good” rows as “bad.” Just like a spam filter, these results are imperfect. But these were out-of-the-gate results! The heuristics and manual exclusions were the product of 10 years of human learning, trial-and-error, and reactive data correction. The learner found 90% of bad rows from a blank slate and an afternoon of human activity.

This experiment showed that a human armed with an AI/ML-based rule learner could boost productivity and data accuracy. These improvements will also preemptively mitigate operational problems resulting from bad demand transaction data, such as bad forecasts or excessive safety stock. And this exact process is easily extended to other areas that challenge supply chain leaders, including:

  • Replenishment order approvals
  • Forecast approvals
  • Replenishment order exclusions
  • Supply transaction exclusions
  • Validating item master data

By reviewing only a relatively small subset of data, we can validate a much larger data set, such as item master records, to see if pallet or weight data may be missing or inaccurate, as is often the case with new products. And these ML-based learners will grow increasingly accurate as they observe more human decisions and thus learn more about the data.

Become More Accurate in Predicting Lead Times

Lead times have been very volatile for the last several years, and GAINS Labs is experimenting with ML-based lead time predictions. In a world with large order sizes, individual item lead times are rarely observed. Combined with lead time volatility, that makes classic lead time estimates unreliable. We experimented with an ML routine to more accurately predict lead time.

We tested our model on a large set of replenishment orders and compared the predictions to the production systems’ alternative lead time estimate. In other words, we asked, “is this ML prediction better than what would have otherwise been used in production?” We weren’t aiming for perfection; we were aiming to beat the status quo.

In our test, the ML technique yielded a better numeric lead time prediction for 73% of out-of-sample replenishment orders. We then layered an ML “meta model” on top of the ML numeric prediction, this time predicting whether the ML numeric prediction or the baseline alternative would be closer to the actual lead time. The meta-model correctly found more than 90% of the numeric model’s “wins” with virtually no errors. The sequence of the two models improved on almost 65% of lead time predictions with no collateral damage.

Interrogating the ML models, the selected predictive features hinted at the sources of the model’s success. These features included:

  • The percentage of vendor orders filled completely vs. partially
  • The count of open orders for a vendor at replenishment order create date
  • The longest observed lead time
  • The latitude and longitude of the supplier

From hundreds of potential predictors and no a priori knowledge of what might predict the lead time, the ML methods independently learned which features are most predictive of lead time. The percentage of vendor orders filled completely in this dataset was the most predictive in this dataset. Thus, the ML model considers past vendor performance when predicting future lead time.

The ML model also identified that latitude and longitude were predictive of lead time, effectively identifying regions within the Americas from which lead times were systematically long/short. Importantly, the model didn’t need predefined regions: it independently discovered meaningful regions from numeric representations of geography: latitude and longitude.

With this data, planners and buyers can make more informed decisions on when to execute replenishment orders and keep their inventory levels aligned.

These are just two areas where GAINS Labs has used ML to produce meaningful results for GAINS customers. The GAINS Labs data scientists employ the latest AI and ML methods to define, test, and deliver new capabilities suited to a specific customer’s needs.

To learn more about GAINS Labs, the collaborative research organization within GAINS focused on developing forward-thinking solutions using data science teams to address clients’ real-world supply chain challenges. Please visit