How TIM Works
After reading through the Introduction to TIM, it is clear what TIM entails. Users might however still have questions regarding the mathematics that support TIM under the surface. This section will provide more information about these underlying mathematical components.
TIM strives to follow all the best practices in time series modeling to achieve the best possible accuracy as fast as possible. More information about the motivation for TIM's architectural design can be found under General Introduction to Machine Learning > Time Series Forecasting. This architectural design is illustrated in the image below.
Some of the specifics of this architectural design are discussed below.
Data Recommendation Template¶
TIM acts like a general time series model building platform. Through Tangent Works' experience in in this domain of time series modeling additional knowledge has been collected from various use cases, indicating which variables should ideally be included alongside the target variable to achieve the highest possible accuracy. In some cases, the data can even be enhanced with various meteorological predictors from different GPS sites. This total knowledge package, personalised to the user's specific challenge, is called a Data Recommendation Template.
Expansion and Reduction¶
TIM is designed to go through all the variables a user includes in the dataset and understand which of these variables are important and which variables do not contribute to the final results. Therefore, adding additional variables to a dataset can never have a negative influence on the achieved accuracy. TIM even generates copious amounts of new features from the original dataset which might be useful for model building, a process called Expansion. If these features turn out to be useless however, TIM can recognise this without compromising on accuracy. After expansion, TIM reduces this newly created set of features to a small, useful subset. This process is called Reduction and is done to achieve model stability and to prevent overfitting.
As explained before, TIM can create many models under one Model Zoo to address different complexity requirements and data availability situations in different usage scenarios. In order to ensure that this is done in a computationally efficient way, an architectural layer is created that exploits similarities in different models and modeling situations to avoid redundant computations. In production, the Model Zoo is able to dispatch the most appropriate model for any situation without the need for a user to be aware of these underlying mechanics.
In anomaly detection, TIM goes beyond the architecture shown above. TIM initially uses its forecasting capabilities to forecast what should be happening in a given situation, following the architecture and steps discussed above. After this, an anomaly detection layer assesses the deviation of the actual situation from this forecast, i.e. the deviation of reality from normality.