Generative AI for Predictive Analytics

Generative AI Use-cases

Enterprise AI/ML use-cases today can be broadly categorized into 3 fields: Natural Language Processing (NLP), Computer Vision/Image Recognition and Predictive Analytics.

Generative AI (Gen AI) capabilities in the fields of NLP and Computer Vision/Image Recognition are well-defined. It all started with ChatGPT [1], which can be considered as the NLP application of Gen AI on textual data, with the underlying pre-trained Large Language Models (LLMs) powering NLP tasks, e.g.

• Question-Answering (QA)/Chatbots
• Text extraction and summarization
• Auto-correct
• Translation
• Classification
• Natural Language Generation (NLG)

The Image Processing/Computer Vision equivalent of ChatGPT would be Stable Diffusion ’s text-to-image Deep Learning model, which people were (or are still) using to generate new art ‘influenced’ by styles of famous artists. This was superseded by the release of GPT-4 which can be considered as a multi-modal model composing [2] text, image, video and speech processing capabilities.

This brings us to the topic of applying Generative AI to Predictive Analytics. The field is not well defined, and we explore what Gen AI can mean in a Predictive Analytics context in the rest of this article.

Generative AI in the context of Predictive Analytics

Predictive Analytics is a powerful paradigm that has widespread applicability in multiple domains: Marketing (Demand forecasting, Churn Prediction, Recommendations), Finance (Dynamic Pricing), Supply Chain Optimisation, Manufacturing (Predictive Maintenance), etc.

Given this wide applicability and the hype of associating Gen AI with everything these days, there are a lot of promises also being made around Gen AI based Predictive Analytics.

The promise here is that you can provide any (complex) prediction problem to a Gen AI Agent, which can then magically process it over relevant enterprise data; and return the perfect prediction with a high level of accuracy.

Gen AI today can potentially provide the below 3 capabilities in the context of Predictive Analytics:

• Synthetic Data Generation
• Natural Language Interface for Optimization
• Transformer architectures to improve Forecasts

Synthetic Data Generation

Given the ‘generative’ nature of Generative AI, Synthetic Data Generation can be considered as the classic use-case of applying Gen AI to Predictive Analytics.

The availability of good quality data (in significant volumes) remains a concern for Supervised ML projects. Synthetic data generation can help by providing high quality data that closely resembles the original data — by implicitly learning the probability distribution of the underlying data.

Synthetic data generation is not new to Gen AI, and Generative Adversarial Networks (GANs) have for long proven quite effective in generating good quality synthetic data.

Intuitively, a GAN can be considered as a game between two networks: A Generator network and a second Classifier network. A Classifier can, e.g., be a Convolutional Neural Network (CNN) based image classification network; distinguishing samples as either coming from the actual distribution or from the Generator. Every time the Classifier is able to tell a fake image, i.e. it notices a difference between the two distributions; the Generator adjusts its parameters accordingly. At the end (in theory), the Classifier will be unable to distinguish, implying the Generator is then able to reproduce the original data set.

For example, TimeGAN is a good example of using GANs model to augment time series data.

Natural language Interface for Optimization

Let us try to understand the steps involved in providing a Natural Language Interface for Planning & Optimization. Given a problem statement in natural language, e.g., “What is the best spend mix for $x budget this year?”

1. Natural Language Understanding (NLU): Understand the problem statement and map it to an Optimization/Planning task.
2. Code Generation (CG): Generate the optimization (model) code corresponding to the task.
3. Execution Engine (EE): Run the code over enterprise data to determine the optimal plan.
4. Natural Language Generation (NLG): Transform the optimal plan/result back into a natural language storyline to be returned to the user.

Typically, the above process would involve running optimization models on data from Sales, Marketing, Distribution and Production — taking into account factors such as Customer demand, Competitor pricing, and Market conditions — to generate an optimal pricing strategy.

Here, we have made a simplifying assumption that relevant source data is readily available in an integrated / modeled fashion. This usually requires significant data engineering efforts in making the data ready to be used by the optimization model. This also requires a certain understanding of the underlying enterprise data, which is easier said than done and we have previously considered the problem in detail in a Conversational BI — SQL mapping context [3].

The above process is not new to Generative AI, and has long existed in the form of Statistical ML models performing Enterprise Planning.

Generative AI can be considered as brining additional capabilities in the form of a Natural Language Interface and Automated Modeling of Complex Scenarios.

• Natural Language Interface: The ability to converse in natural language, both in terms of specifying the problem and comprehending the result.
• Automated Modeling: The second differentiator is in terms of Gen AI’s ability to process large amounts of data over multiple scenarios and models to determine the best options for cost savings and operational efficiency across the enterprise.

This significantly lowers the barrier to entry making them more accessible to business users.

Gen AI Forecasting—Prediction Accuracy

Predictive Analytics involves analyzing large amounts of historical data, incorporating factors such as seasonality, trends, and economic conditions; to train a Machine Learning/Deep Learning model that can make accurate predictions / forecasts.

The key question here is if Generative AI can improve the Prediction accuracy?

This is easier said than done as there is no consensus on what is a Generative AI model for Predictive Analytics.

Even if we consider Transformer based Predictive Models as the closest to Gen AI applied to Predictive Analytics, it is very difficult to argue that it would automatically lead to a better prediction accuracy in all scenarios.

Currently, forecasting at most enterprises relies on generic forecasting models and ad-hoc (manually generated) statistical models based on human expertise, often applied directly on Excel data. Generic forecasting models use standard algorithms based on the assumption that demand can be predicted uniformly, irrespective of their domains and geographies. This one-size-fits-all approach yields a prediction that fails to exploit the different and diverse demand drivers, market conditions and evolving consumer behaviours.

With respect to the applicability of Deep Learning (DL) based approaches for Prediction problems, the jury is still out. Some studies [4] claim that DL based approaches provide significantly better results than Statistical ML models. Given the time-series nature of most forecasting problems, it is not hard to imagine the applicability of Recurrent Neural Networks (RNNs) / Long Short-Term Memory (LSTMs) for forecasting. At the same time, studies [5] have shown that Statistical models, e.g., Exponential Triple Smoothing (ETS) and AutoRegressive Integrated Moving Averages (ARIMA) and Support Vector Machines (SVM); still outperform DL based forecasting solutions.

The conclusion so far is that a hybrid (ensemble) Statistical and DL architecture is the best, with the statistical model providing the base forecast based on historical (predictable) patterns, and the DL part complementing to predict the spikes based on external factors.

Let us dive a bit deeper into this complex world of forecasting. The fundamental reason why DL has been so successful with images and text is that the main signal is usually inside the analyzed object — we may be processing or classifying images, or correcting typos in the text, but in all these cases we work with the objects (images, text) themselves.

In time series data, on the other hand, the time series itself usually bears only part of the signal, while a significant portion of the factors influencing the forecast are external to it.

For example, product demand is influenced by client behavior, external economic factors, competitor activities etc. If we consider those factors constant, using just the time series data is sufficient for the forecasting. But in practice, such factors are changeable, and in order to have timely high-quality forecast, we need to have those factors embedded in the model. For example, if client behavior (e.g. traffic in the stores) changes due to some external reasons (e.g. unexpected bad weather), we need to have that information embedded in the model (e.g. by incorporating weather as factor in the model), so that we can have a quality forecast before the actual demand changes.

Thus, whenever we need to account for external factors and incorporate them in the model, we need to look for approaches that work best with huge sets of different types of data — and DL by itself is not sufficient. Studies such as [6] have shown that Transformer architectures [7] can outperform RNNs / LSTMs in such scenarios. Transformers eliminate recurrence, thus decreasing complexity, and enabling parallelization — leading to overall faster computation and greater accuracy for time series forecasting. [8] is an interesting work in this direction that uses Bidirectional Variational Auto-Encoder (BVAE) to first augment the input time series data, and then applies denoising and disentanglement on the generated data to improve accuracy and interpretability.

To conclude, ensembles of models is the way to go, and GPT-4, long argued to an ensemble model; can actually lead to better prediction accuracy.

However, a comprehensive study is needed to validate this and the improvement may not be evident on all datasets. So there is no need to rush to GPT-4 for all your prediction problems, at least not yet?! Generative AI can also act in an Auto ML [9] fashion trying out different combinations of configurable hyperparameters, e.g. number of hidden layers, number of neurons per layer, the activation function, optimizer, batch size, number of training epochs, etc. — to identify the optimal neural network architecture for a given problem and dataset.

References

1. D. Biswas. ChatGPT internals, and its implications for Enterprise AI. https://medium.datadriveninvestor.com/will-chatgpt-disrupt-enterprise-ai-7b83b7591c1e
2. D. Biswas. Compositional AI: The Future of Enterprise AI. https://towardsdatascience.com/compositional-ai-the-future-of-enterprise-ai-3d5289dfa888
3. D. Biswas. Conversational BI: Text to SQL. https://towardsdatascience.com/conversational-bi-text-to-sql-c9f52a89acc5
4. F.M. Bianchi, et. al. An overview and comparative analysis of Recurrent Neural Networks for Short Term Load Forecasting, 2017. https://arxiv.org/abs/1705.04378
5. S. Makridakis, et. al. Statistical and Machine Learning forecasting methods: Concerns and ways forward, 2018. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5870978/
6. Nino Steven. Transformers and Time Series Forecasting, 2020. https://arks.princeton.edu/ark:/88435/dsp01kk91fp583
7. Niels Rogge, Kashif Rasul. Probabilistic Time Series Forecasting with Transformers. https://huggingface.co/blog/time-series-transformers
8. Y. Li, et. al. Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement, 2023. https://arxiv.org/abs/2301.03028
9. D. Biswas. Is AutoML ready for Business? https://towardsdatascience.com/is-automl-ready-for-business-ffe4c2d8b5af