What are the most common reasons for an ML model to fail in production?

I think that the main problem is usage of "wrong" data to create ML models. In almost all papers researcher used academically collected data - data collected very carefully for this study. Such data are very good for work but unfortunately they have no common with routinly collected data. The last one have huge number of missing values, collected in different way and so on. AS a result, ML system prepared for academic data work poorly on routinly collected ones.

Data quality is one of the, if not the, most critical aspects while building an ML model. Too often I have seen ( and may have been guilty myself) of worrying about the quality and size of data in training but not paying attention to the actual workings of gathering that same data in the real world or what we call production. One of the most common mistakes I have seen is folks using claims data without accounting for the data lag in the real world. The most recent 30-45 days insights become irrelevant more often than not on claims only models and thus features or insights that looked like game changers in training and even testing suddenly become unusable in production and the model breaks apart.

Data quality is one of the most important factor that drives the outcome of any ML model. While we thrive to maintain high quatlity of data, often times we miss on the precursors like, the basic understanding of the raw data; the way it has been sampled; scope of collection; methods used in collection/validation/cleansing; and so on. That leads to - incorrect interpretation of outliers; loss of valuable data; bias; etc., and that leads to an inacurrate model.

This problem is commonly known as domain mis-match; When training data vs testing data (production) has different distribution, the model trained would perform poorly In testing. To solve this problem, domain adaptation is applied using relevant small amount of test data similar to Production use case.

Imagine your ML model as a chef, and data quality as the ingredients. If the ingredients are bad, your dish won't taste good. Similarly, if your data is of poor quality, your ML model won't perform well. Data quality means having accurate, complete, consistent, and relevant data. It's like making sure all your recipe ingredients are fresh, measured correctly, and suitable for your dish. If your data isn't top-notch, your ML model might make biased or inaccurate predictions. It's like expecting a recipe to turn out perfectly with bad ingredients. Poor data quality can lead to prediction errors. Imagine serving a dish with a critical ingredient missing – disaster! So, follow best practices for collecting, cleaning, validating & labeling data.

O desvio de dados é um problema frequente que pode levar à falha de modelos em produção (quando fazemos o deploy). Isso ocorre quando a distribuição ou as características dos dados que o modelo recebe em produção mudam ao longo do tempo. Essas mudanças podem ser causadas por diversos fatores, uma em questão que mais me ocorre é a Sazonalidade: Mudanças repentinas nos dados e o modelo acaba não reconhecendo os padrões.

When the properties or distribution of the data that the model receives in production vary over time, it's known as data drift. These changes can be caused by a variety of variables, including user behavior, market trends, seasonality, and external events. Data distribution fluctuates with time. Model accuracy may be impacted by this change. To make sure the model continues to be useful in reflecting the underlying relationships, regularly review and update it to take into account modifications in data patterns.

Data drift is one of the main reason why your model is not working well in practice. Because it is impractical to incorporate all the real world instances of a category into a training set, often we subsample certain portion of it that characteristically represents a concept of that category. This results in lack of generalization capability of the model to instances outside of the training distribution, which degrades the model's performance. For instance, a cow is often captured eating grass so significant portion of the cow image samples would be centered around the grass background. However, what if a cow is walking on the beach? Can your model recognize it as accurately as it did before for the grass one? The answer is maybe or No.

Detecting trends and shifts in data promptly can increase the robustness of the ML model. Given the rapid changes post-Covid, staying ahead is critical. When shifts haven't yet significantly changed distributions, including features to reflect these recent trends can provide the model with an advantage. For instance, features like the year-over-year percentage change in user behavior can strengthen the model's performance. However, trend features should not replace the need for regular retraining.

Another common mistake is the use of ‘future’ information. It is important to check that the data used for training contains only the information that was available at that time, in particular when there are features that change status and so on.

Imagine your ML model as a skilled driver navigating through a city. Initially, it learns the routes and traffic patterns, much like a GPS system. However, due to events like city development or road closures, the usual routes might change over time, introducing concept drift. Now the cityscape would have changed. To overcome this, you can implement adaptive learning techniques and algorithms. Think of it as giving your driver a real-time traffic update system. Online or Mini-Batch learning can be employed to update the model as new data arrives. Specialized algorithms like CVFDT (Concept-adapting Very Fast Decision Trees) and DDM-EDS (Google it😝) are designed to handle concept drift by continuously adapting the model to changing data.

ML models failing in production due to data drift is a prevalent issue. - 📉 Data drift arises when the incoming data to the model deviates from the training data, impacting model performance. Factors like seasonality, user behavior, market trends, and external events contribute to this shift. - 🔄 To mitigate data drift, implement monitoring methods 🕵️, utilize metrics to compare data statistics, and establish thresholds for acceptable drift. - 📊 Regularly update or retrain the model based on observed drift to maintain accuracy and relevance. -⚙️ Failure to address data drift can lead to outdated models and compromised predictions. -🚀 Stay vigilant to ensure ongoing model effectiveness.

Concept drift requires rebuilding the model either fully or incrementally. It does not occur randomly, so one has to follow the environment in which the models are deployed to anticipate and prepare for a concept change. Usually, a good communication channel and change management processes can effectively handle concept drift ahead of performance impacts.

Concept drift also tends to crop up when the problem you’re solving isn’t well-understood from the beginning. That’s why it’s important to periodically revise your problem statement as you gain more context and insight. This step will help input/output align with what you seek to accomplish over time.

ML models are not always successful in production, and one of the main reasons for this is the misalignment between actual business needs and ML objectives. Misalignment occurs when the problem that the ML model is trying to solve is not well-defined, relevant, or measurable in terms of the desired business outcomes. For example, if a company wants to increase customer retention, but the ML model is trained to predict customer satisfaction, there may be a mismatch between what the model optimizes and what the business cares about. It is therefore important to start with a clear and specific problem statement that aligns with the business goals, and then translate it into a machine learning objective that can be quantified and evaluated.

Model complexity can lead to ML model failures in production due to overfitting, poor generalization, and difficulties in interpretability. Overfitting occurs when a model captures noise instead of underlying patterns, affecting its performance on new data. Techniques like regularization, pruning, and feature selection help address overfitting. Model interpretability is crucial for understanding and trusting model decisions. Model compression optimizes computational efficiency, aiding deployment. Balancing complexity with metrics and frameworks ensures model quality and trade-offs. Regularization, pruning, feature selection, and compression mitigate risks associated with overly complex models.

Traditionally, model complexity in machine learning has been about the trade-off between complexity and generalization. This will continue to be relevant. However, with the advent of large language models, another dimension of complexity has arisen. One that is related to cost. It may seem easy to just throw a problem at a large language model and receive answers. People are racing to show off their skills, sometimes not considering simpler solutions. Using large language models for classification, regression or even feature extraction tasks is an overkill - a costly one not only in monetary terms, but also in its' impact on environment.

Optimal model complexity is crucial for ML success. Overly complex models risk overfitting, learning noise instead of patterns, hindering generalization. Balancing complexity through regularization, pruning, and feature selection ensures interpretability, maintainability, and robustness, enhancing model usability and trustworthiness while mitigating issues in production.

The point about model complexity aligns with the challenges I've faced in deploying ML models. It's crucial to strike a balance between complexity and interpretability, especially in sectors like finance or healthcare where explainability is as important as accuracy. Techniques like LIME (Local Interpretable Model-agnostic Explanations) or SHAP (SHapley Additive exPlanations) can be important tools for making complex models more transparent and trustworthy to end-users.

ML models have become the go-to solution for a wide variety of real-world problems. We are exposed to an overabundance of data, much of which is either irrelevant or needs several transformations. Training a model on a large number of parameters and unclean data often make it complex. Choosing the relevant attributes and parameters is a must. Moreover, the data must be standardized and normalized to suit the context. Techniques such as the assessment of correlation between dimensions, dimensionality reduction and cross-validation can come in handy. The model should be designed in such a way that its predictions remain traceable and the underpinning modus operandi remains understandable. Both overfitting and underfitting should be avoided.

Failed productionized ML models arise from sudden changes in the input data, inadequate integration with other systems and retraining/versioning frequency. MLOps is the newest area addressing these issues through rigorous and automated monitoring including report logging across the entire pipeline from reading, processing and cleansing the data, to (re)training, making predictions, evaluating model performance, distributing the model output into other systems, logging model and data versions. Another angle not often considered directly is platform cost management which can alter the scalability of high frequency ML retraining. Expanding MLOps to incorporate cost monitoring is an increasing need for successful ML production deployment.

Operational issues have become so prevalent that they've given birth to new disciplines like MLOps and DataOps. Books and articles written about it, where tech companies try to generalize their solutions into best practices, processes, and frameworks. However, for larger tech companies it might make sense to strive for a bespoke solution that would fit their unique infrastructure and in-house built tech stack. For other, non-technology companies, this might not be equally prudent for every single company to follow suit. Because of this, a proper foundation needs to be built - a data platform that will bring together not only storage and data processing technologies but also tools and processes for ensuring (and enforcing) data quality.

ML models can face operational challenges such as deployment issues, latency problems, or scalability constraints. A model that works well in a controlled environment may fail in production due to unforeseen operational issues. Regularly monitor and maintain the production environment, ensuring seamless integration and consistent performance.

In deploying ML models, focus on integration with existing systems and robust pipelines for data validation, model versioning, and automated retraining. Monitoring is essential—implement systems to track performance and detect data drift or accuracy degradation. Continuously evaluate predictions against real-world outcomes to enhance adaptability. Prioritize security and compliance, adhering to data privacy laws and encrypting sensitive information. Plan for scalability using cloud solutions or scalable infrastructure to handle variable workloads. Success in ML deployment hinges on integration, performance, adaptability, and real-world applicability, beyond just model accuracy.

Continuous monitoring and robust monitoring framework is key to avoid performance degradation and concept drift over the period of time in a scaled deployment environment.

Trust in ML model's output is not only impacted by the obvious - quality of predictions. Factors ranging from explainability issues, human inability to relinquish control, overconfidence in "guts" or intuition play equal if not bigger role. Add to that aspects like fear of job security, lack of willingness (or inability) to adapt or even some hidden agendas, and we have several issues to solve. What makes things worse, being non-technical in nature, they are happily ignored as they fall beyond the engineers' comfort zone. This is exactly why those issues should be planned for and addressed proactively by management.

Human factors play a pivotal role in the success or failure of ML models in production. Even if a model is technically robust, its effectiveness depends on addressing human aspects. Factors like user expectations, perceptions, preferences, emotions, and behaviors can significantly impact the model's acceptance. To mitigate these risks, involve users and stakeholders throughout the development and deployment process. Utilize user research, design thinking, and usability testing to align the model with user needs and values. Incorporate ethical AI principles to ensure fairness and transparency. Neglecting human factors can lead to user dissatisfaction, confusion, and frustration, undermining the model's overall success. 🤖👥🔍

I would add Human Factors. Humans, who would eventually, implement these ML-based models into production suffer from cognitive biases. Unless we account for these biases in the ML models, there will always be a difference between expectation and reality. For instance, while building a forecasting algorithm through Multiple Linear Regression, human beings often ignore some critical variables or may stick to some other depending on what their bosses want them to do. As such, the models' outcomes suffer.

In his book "Designing Agentive Technology", Chris Noessel talks about the types of tasks that humans are bad at and AI is good at. Understanding how humans can interact with AI and designing suitable control mechanisms for handoff and handling errors are among the few important points to consider. This is especially important with the rapidly developing AI regulation space.

Human factors, including biased training data, inadequate domain knowledge, or misinterpretation of model outputs, can contribute to failures. Biases in the training data may result in discriminatory predictions, while insufficient understanding of the domain can lead to the model making inappropriate decisions. Involve domain experts, conduct thorough analyses, and address biases to mitigate human-related risks.

Consider the impact of cultural diversity on user interactions with the ML model. Cultural nuances can influence how users perceive and interpret model outputs, affecting its overall effectiveness. For instance, certain features or recommendations may be well-received in one culture but misunderstood or rejected in another. Conduct cross-cultural usability testing and incorporate diverse perspectives in the development process. Additionally, share user success stories to highlight the model's positive impact and build trust. Regularly gather feedback post-deployment to adapt the model to evolving user needs. 🌍🤝

In my experience, the most common reason a machine learning application fails is that it doesn't actually solve any problems. In other words, it's not actionable. The most common scenario is with predicting customer churn, where the model either picks up on focus on factors we can't influence, rendering their insights impractical. Or the model identifies churn risks too late, making interventions difficult.

La calidad de los datos desempeña un papel fundamental en el desarrollo y rendimiento de modelos de machine learning. Las relaciones causales presentes en los datos pueden evolucionar con el tiempo debido a factores externos no controlados por el modelo. Por lo tanto, es imperativo realizar un monitoreo constante de los indicadores de desempeño del modelo y ajustar los hiperparámetros cuando el escenario cambia. Este proceso de calibración continua es esencial para garantizar que el modelo se adapte de manera efectiva a los cambios en los datos y mantenga su capacidad predictiva a lo largo del tiempo.

A model is as good as the data on which its trained on. The performance of a model in production may vary in the course of time coz of various reasons like data drift, concept drift, etc. Its very important to monitor the performance of models in production and have feedback systems for the models to relearn. So, without a proper feedback system for models in production will lead to model failure in production.

To take deep learning from research to production, performance needs to be qualified by the latency and heaviness of the model. Indeed, whether we are talking about AI at the edge or deployed in the cloud, it is always preferable to have a small memory footprint and short execution time.