Trendalyze

The way that Trendalyze works is that you inspect time series data looking for micro-trends (motifs) that are of interest. Take predictive maintenance as an example. What you are trying to achieve is to predict when a particular asset is likely to fail and repair it before it does so. To do this you require a large body of historical time series data. Instead of developing a machine learning model, with Trendalyze the software – using built-in machine learning algorithms – will identify specific micro-trends in the time-series data that can be used to make these predictions: once a relevant motif has been found you can use this as your predictor. And, specifically, while it may be relatively easy to spot a pattern – when displayed visually – it is not easy to spot a recurrence of that pattern, let alone any slight variation on that pattern. This is what Trendalyze aims to automate for you. The software applies search principles to look for instances of an identified motif within the dataset being examined. Once discovered you can set threshold parameters around your motif (in other words, supporting variations to the motif), which can be used to predict assets that need maintenance.

Fig 03 The Trendalyze motif library

Meaningful or known motifs can be stored in a library – see Figure 3 – which can then be used to search for similar cases in different data sets or to build predictive applications. Specifically, you can set the software to look for any particular motif or pattern of motifs that and raise an alert or other action when it occurs. Conversely, you can automate an appropriate action when a motif doesn’t occur, and you can combine motifs to provide correlated search. For example, in anti-money laundering you might look for, say, five anomalies occurring within a given time period. Note that you can still create predictive models (via the product’s support for Spark and TensorFlow) if you wish to, though the reason that Trendalyze provides its own modelling capabilities is to give users the ability to deploy machine learning without having to understand modelling and without needing expensive data scientists.

Experience suggests that motif-based prediction can be much more accurate than using conventional approaches based on data science. For example, at Lockheed Martin Trendalyze was able to reach a 97% success rate at predicting engine failures, while other approaches that were tried got no better than 50%. In the pull-out quote from RummyCircle.com, which is a gaming site, a similar dramatic improvement (actually we understand that 75% is a conservative estimate) was performed but here the “assets” are people and the predictions are about churn, based on how much gamblers are spending over a relevant period of time.

The other big advantage, apart from accuracy, is that you don’t need any coding to achieve these results. In the Lockheed Martin example there was no coding required to get 97% accuracy, but alternate approaches did require coding by data scientists. Thus, there is significant cost saving potential resulting from the deployment of Trendalyze.

The Bottom Line

We are impressed by Trendalyze. One of the big problems with artificial intelligence is that it is often difficult to put machine (let alone deep) learning into operation, typically because there is a disconnect between data scientists, who often work for business departments, and IT. As an end user tool Trendalyze obviates this issue.