<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by aestew on Medium]]></title>
        <description><![CDATA[Stories by aestew on Medium]]></description>
        <link>https://medium.com/@aestew?source=rss-69a1a28d2b06------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ADwd5vi4xo0gQW099zcsuA.jpeg</url>
            <title>Stories by aestew on Medium</title>
            <link>https://medium.com/@aestew?source=rss-69a1a28d2b06------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 11:19:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aestew/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Feature Tokenizer Transformer Model Performance in Predicting Flight Delays]]></title>
            <link>https://medium.com/@aestew/feature-tokenizer-transformer-model-performance-in-predicting-flight-delays-96126989e133?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/96126989e133</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[flightdelays]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Sun, 14 Dec 2025 21:37:08 GMT</pubDate>
            <atom:updated>2026-01-16T00:25:37.033Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://github.com/aestew/flight_delay_predictions/tree/main"><em>Github: flight_delay_predictions</em></a></p><h3>Overview</h3><p>75 million rows of DOT flight delay data and 132 million rows of weather data from NOAA were joined using Apache Spark and analyzed to evaluate flight delay predictions under significant class imbalance and temporal distribution drift. Flight delay predictions require modeling complex interactions, including flight schedule, airline, flight network effects, temporal dependencies and external issues including maintenance, staffing and security. Here, we use three model types, graph and temporal features to compare predictions addressing imbalance through downsampling and focal loss.</p><p>We evaluated three model families: Logistic Regression, Tree (XGBoost and Decision Tree) and two neural models were considered (MLP and FT-Transformer). F2 and model generalization were both assessed as well as, Recall, Precision, PR-AUC.</p><p>The prediction task was defined as whether a flight would arrive more than 15 minutes late based solely on information available one hour before takeoff, ensuring a realistic operational inference setting. Technical Details and outcomes are below.</p><h3><strong>Distributed Data Engineering and Temporal Join Optimization</strong></h3><p>The complexity of joining 74M flight records with 132M weather observations across temporal boundaries required sophisticated distributed computing strategies. We implemented Apache Spark’s RepartitionByRange optimization to ensure data locality for temporal joins, partitioning both datasets by timestamp ranges to minimize cross-partition shuffling during the nearest-neighbor weather station matching process. This approach was critical given that each flight required identification of the closest weather observation within a 1-hour window prior to scheduled departure — a operation that would be computationally prohibitive without proper data partitioning. The RepartitionByRange strategy reduced join execution time from several hours to under 45 minutes by co-locating temporally adjacent records on the same cluster nodes, enabling efficient range-based lookups for weather conditions. Additionally, we applied secondary partitioning by airport codes within each temporal range, further optimizing the geographic nearest-neighbor searches required to match flights with appropriate weather stations. This distributed architecture proved essential for handling the scale and complexity of multi-modal temporal data integration while maintaining strict constraints on data leakage and operational realism.</p><h3>Advanced Feature Engineering and Graph Networks</h3><p>To capture both systemic and temporal delay dynamics, we engineered features including graph-based features (PageRank) for airport and lag features identifying previous flight arrival. This graph-based approach quantifies how delays at hub airports (high PageRank scores) disproportionately impact downstream flights compared to delays at smaller regional airports. The temporal lag features incorporate previous flight arrival information, but only when the actual arrival occurred at least 1 hour before the subsequent flight departure — ensuring no data leakage while maintaining operational realism.</p><h3>Addressing Class Imbalance</h3><p>The severe class imbalance (approximately 20% delayed flights) created significant training challenges for traditional approaches. Downsampling methods, while improving recall for tree-based models, systematically inflated delay predictions and caused precision collapse during validation. Our FT-Transformer implementation leveraged Focal Loss with parameters α=0.25 and γ=2, which dynamically reweights difficult minority-class examples during optimization without distorting the underlying class distribution.</p><p>Mathematically, Focal Loss modifies standard cross-entropy by adding a modulating factor (1-pt)^γ that down-weights easy examples and focuses learning on hard negatives. This approach proved crucial for temporal stability — while tree-based models showed PR-AUC degradation of up to 0.27 across time splits, the FT-Transformer maintained stable performance (PR-AUC 0.438–0.452, Δ=0.013) across rolling validation windows, demonstrating superior generalization under temporal distribution drift.</p><h3>Production Scale and Infrastructure: Transformer Training</h3><p>The distributed processing pipeline handled 200M+ records using optimized Spark joins and GPU acceleration. Initial CPU-based transformer training required 15+ hours per epoch; migration to NVIDIA T4 GPUs reduced training time to approximately 4 hours while enabling larger batch sizes and more sophisticated hyperparameter searches. The production deployment constraints (1-hour prediction window) required careful temporal validation using rolling windows that simulate real-world operational scenarios where models must predict delays for future flights based only on historical patterns.</p><h3>Results</h3><p>Results of ML model experiments are below and highlight the limitations of data-level imbalance corrections for tree-based models in non-stationary settings and demonstrate the advantages of loss-based approaches, such as focal loss, when combined with expressive neural architectures for large-scale, temporally evolving prediction tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZTJ8FqhjmL7EQkg9yTJMnw.png" /></figure><p><a href="https://github.com/aestew/flight_delay_predictions/tree/main"><em>Github: flight_delay_predictions</em></a></p><p><em>Notes and contributions:</em></p><p><em>Initial pipeline build and XGBoost was built and tuned by </em><a href="https://www.linkedin.com/in/kristen-lin/"><em>Kristen Lin</em></a><em>, Random Forest by </em><a href="https://www.linkedin.com/in/chadvo/"><em>Chad Vo</em></a><em>, logistic regression by </em><a href="https://www.linkedin.com/in/priscilla-siow/"><em>Priscilla Siow </em></a><em>and Decision Tree by </em><a href="https://www.linkedin.com/in/uma-krishnan/"><em>Uma Krishnan</em></a><em>. </em><a href="https://www.linkedin.com/in/amyesteward/"><em>I</em></a><em> built, tuned and tested MLP model and FT-Transformer model, supported on pipeline build and created graph and temporal features.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96126989e133" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Extracting Scalability and Value in your Data Team: Foundations]]></title>
            <link>https://medium.com/@aestew/extracting-scalability-and-value-in-your-data-team-foundations-21966cd90946?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/21966cd90946</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Tue, 26 Aug 2025 19:53:52 GMT</pubDate>
            <atom:updated>2025-08-26T19:53:52.124Z</atom:updated>
            <content:encoded><![CDATA[<p>By Amy Steward</p><p>Significant investments into shiny Data Analytics and Data Science teams look great on podcasts. However, churn, confusion and re-work caused by a poor data foundation wastes literal years of work and millions in investments with almost no measurable outcomes. Inefficient companies may rely on Analysts and DS teams to build pipelines without proper version control or change management and no overarching data strategy. This means when someone leaves the company, their pipeline, KPIs, data and work fade with them, only for the next DS to re-invest in reinventing what someone spent years building.</p><p>The key to a solid data foundation is a unified approach that spans data ingestion, transformation, storage, and governance across the entire organization. Workflows should exist to integrate data from all 3rd party sources (e.g. Salesforce, Jira, etc.), new product features, marketing tools (Marketo) and sales tools. Data Engineers should be partnering with product managers and leaders across the company as needs evolve and change to ensure data is accessible and ideally in a data lake (or able to move to a data lake) when needed. Data Strategy documents should outline a minimum structure of how to format the data.</p><p>When working with Product teams, Data Engineers and Data Analysts should be working closely together to identify key outcomes and flows of new products. For every new product, teams should define a clear “success metric” and the workflow that drives it. Data Engineers should ensure that data is collected and stored in the data lake, while Analysts handle initial performance analysis until the product stabilizes — keeping Data Scientists focused on specialized problems.</p><p>A solid DE team will have a close working relationship with engineering, specialize in database and pipeline development and have a solid understanding of KPI development. Close working relationships with Data Analysts and Product Managers is critical when creating pipelines and tables. A Data Strategy should outline standards including how metrics are managed, where data is kept, change management, DBT and/or Git workflows at a minimum.</p><p>A well-oiled and mature data organization requires a solid data foundation. With clear responsibilities outlined, data organizations are happier, and more productive. However, the biggest benefit of running a structured data organization today is significant automation through ML GenAI. With solid pipelines and a trust in the data organization, companies will be able to harness their most valuable asset — data — turning it into measurable revenue streams through responsible and strategic data monetization.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21966cd90946" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Extracting Scalability and Value in your Data Team: Data Science]]></title>
            <link>https://medium.com/@aestew/extracting-scalability-and-value-in-your-data-team-data-science-0c334c812f60?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c334c812f60</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Mon, 25 Aug 2025 03:25:04 GMT</pubDate>
            <atom:updated>2025-08-25T20:46:05.727Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Outcomes of capable data orgs, and some keys to get there: Data Science</em></p><p>By Amy Steward</p><p>Months of work and millions in spend vanish when Data Science projects are shelved. This significant mis-match not only causes a drop in morale, it results in months of wasted time and financial investment. Further, due to their significant technical skills, DS are often spread thin focusing on numerous projects. Here I will review how to not only optimize DS spend, but build a happier and more robust team. Business results from this program include: democratized ML across the organization, thoughtful and data driven product decisions, CS and Sales team optimization, and the ability to build agents to automate workflows.</p><p>Generally, DS should focus on: creating a scalable and repeatable testing platform (including company documentation on acceptable significance), segmenting customers for campaigns (clustering), revenue prediction based on trends and lifecycle modeling through identifying likely churn customers.</p><p>Successful DS organizations should be vertically integrated and closely partnered with all divisions of an organization. Through building camaraderie with a partnered org, successful DS will be able to deeply understand challenges and align to build key critical products and ensure implementation. Baseline analyses should be done first, including clustering and risk identification. Close partnership with leaders will be required to ensure that IC’s fully understand use cases before moving to more complex and nebulous solutions. Success should be measured by relative adoption of a tool, which is an inherent measure of success and efficiency gain (i.e. ROI). All tools should have a planned review window and measurable drift thresholds. When validation requirements aren’t met, models should be retrained or retired.</p><p>Data Scientists (DS) should leverage data curated by Data Analytics and Data Engineering Teams. In some cases, it may make sense for a Data Scientist to dig into log data, however — pipelines remain a Data Engineering function. This allows DS to focus on rapidly prototyping, improving models, partnering with stakeholders and communication with leadership. Successful ML models and agents should be launched and maintained by MLOps teams until expiration or model drift past acceptable validation rates.</p><p>Through focused development and close partnership with stakeholders, Data Scientists can develop trust with an organization to build useful and successful tools. Mature organizations with trusted partnerships should be implementing features including linking technologies such as salesforce and slack with GenAI to create quick customer references, implementing RAG with GenAI to improve communications and developing AI agents. Again, while an organization could quickly implement the former, such tools will not be used by the organization without building partnership and trust.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c334c812f60" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Extracting Scalability and Value in your Data Team: Analytics]]></title>
            <link>https://medium.com/@aestew/extracting-scalability-and-value-in-your-data-team-analytics-4ba4c3fe450e?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ba4c3fe450e</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Sat, 23 Aug 2025 15:20:44 GMT</pubDate>
            <atom:updated>2025-08-23T15:20:44.152Z</atom:updated>
            <content:encoded><![CDATA[<p>By Amy Steward</p><p>The last 10 years have created a focused and exciting environment for data as organizations have heavily invested in data teams. While leaders have invested significantly in this opportunity — there is a large gap between companies on the purpose and goal of this team. In this post, I will specifically focus on analytic functions and outline basic outcomes whereby these teams can maximize impact and optimize company investment.</p><p>Capable analytics teams are plugged into all core drivers of the business — starting with development and maintenance of ‘North Star’ metrics — tied to revenue and customer value. This includes trends and cycles of: sales, revenue, product usage and funnels (focusing revenue driving products). Product value-proposition should be reverse engineered through clustering to understand actual value customers are receiving (e.g. are customers actually using the new AI feature, or do they just do the 1 thing and move on).</p><p>With a solid, scalable data platform and a united data engineering team, data analytics teams should partner with stakeholders to identify critical KPIs. KPIs should be well defined, easily accessible through a core data table or BI tool and consistent across the organization. Weeks per month will be saved by creating a united language. This allows analysts to move onto other projects, instead of re-evaluating and re-calculating KPIs. Outcomes allow leaders to readily access these KPIs and discuss them both within the organization and externally with minimum confusion.</p><p>Critical KPIs should be thoroughly defined in an accessible place for the whole company. This includes basic counts such as: revenue, customer counts, customer status, feature identification, feature purpose and outcome (e.g. what is the thing that needs to be done for engineers and PMs to high-five). This data should be built through a solid and secure pipeline and available for all to analyze.</p><p>Analysts should be experts on this data and the hourly, daily and monthly trends. Analysts will serve as trusted advisors for C-level executives to lean on when building out yearly goals. Through this partnership, reasonable goals can be constructed across the organization allowing for tracking and improving company focus, productivity and morale.</p><p>To extract the most value, data teams should be integrated vertically throughout the entire organization. Data leadership should be focused on company alignment and data clarity while continually advocating for their team. Implementing this framework delivers measurable time savings, more efficient analysis, and a clear runway toward agentic AI development and advanced ML models.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ba4c3fe450e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Extracting Scalability and Value in your Data Team: Introduction]]></title>
            <link>https://medium.com/@aestew/extracting-scalability-and-value-in-your-data-team-introduction-484502d133eb?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/484502d133eb</guid>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Sat, 23 Aug 2025 03:54:47 GMT</pubDate>
            <atom:updated>2025-08-23T03:55:27.657Z</atom:updated>
            <content:encoded><![CDATA[<p>By Amy Steward</p><p>Data is a company’s most valuable asset. From app usage, to messages, to customer sentiment datasets can be curated from information customers are continually creating. This data can drive significant growth — if you know how to use it. In the last year, two massive opportunities have surfaced, which many companies have yet to leverage. This includes leveraging AI Agents and machine learning prediction models and monetizing and leveraging data internally and externally. Today many data programs aspire to leverage these tools, however millions of dollars are lost and significant confusion arises without a cohesive and comprehensive data strategy.</p><p>Well aligned data teams are driven by their data foundation, clear roles and end-to-end asset management. This includes retiring expired KPIs, dashboards and models. For a data org to be effective, data needs to be at the heart of a company’s fabric and tied in at every level of an organization. While the hype-train for data was strong 5 years ago, companies will benefit from reviewing or implementing a data strategy. Without this, advanced data analyses will likely result in significant waste of time, resources, and morale.</p><p>Through thoughtful execution, a well run data team can leverage a clean and scalable data pipeline and provide consistent metrics. This allows stakeholders to self-analyze, actionable and thoughtful KPIs, partner with C-level and Board members to chart out growth areas. Scalability built data science teams should partner with company organizations to implement RAG, integrate 3rd party tools and streamline workflows — saving significant time and money. Further, once a team has a solid data platform, teams can focus on building customer ML models to forecast churn, identify customers by stage, and monitor for onboarding progress and so much more.</p><p>This series will look at outcomes well run data teams can achieve with focused organization, a mature data platform and pipeline, discuss foundations of a data organization, and finally identify team structures and tools available to develop and build this data team. Using these pieces, companies can execute an end-to-end data strategy to leverage their most valuable asset: data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=484502d133eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Predicting Corn Yield in the US]]></title>
            <link>https://medium.com/@aestew/predicting-corn-yield-in-the-us-d69190b65464?source=rss-69a1a28d2b06------2</link>
            <guid isPermaLink="false">https://medium.com/p/d69190b65464</guid>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[uc-berkeley]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[aestew]]></dc:creator>
            <pubDate>Mon, 18 Aug 2025 01:57:07 GMT</pubDate>
            <atom:updated>2026-01-16T00:14:28.078Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Amy Steward, Chad Adelman, Michael White and Yoko Morishita</strong></p><p>Corn yields in the U.S. are increasingly volatile due to climate change, but combining environmental, genetic, and geographic data can deliver accurate forecasts — our model reached an RMSE of 18.02 bu/ac, rivaling top recent studies.</p><p>Corn yields in the United States are becoming more unpredictable due to climate change and shifting weather patterns, creating challenges for farmers, policymakers, and the broader agricultural economy. Accurate forecasts help farmers plan planting and resource allocation, while enabling policymakers to anticipate supply fluctuations and safeguard food security. Our work focused on predicting yields across seven Midwestern states (Iowa, Illinois, Nebraska, Missouri, Minnesota, Indiana, and Ohio) using a rich dataset that combined historical yield data from the USDA, climate variables from NOAA, and adoption rates of genetically engineered crops. Analysis revealed that certain climate conditions — particularly summer rainfall and high temperatures during July and August — have strong correlations with yield outcomes, and that genetic modification features, especially the percentage of stacked genes, are among the most predictive factors. While our models ultimately relied on machine learning to produce forecasts, the underlying insight is that combining environmental, genetic, and geographic information can improve the ability to anticipate and respond to the volatility of modern corn production.</p><h3>Initial Model Builds</h3><p>For our baseline model, we used the average yield in bushels / acre and compared subsequent models to this. Our goal for experimentation and model building was to have &lt;1.25 MAE generalization ratio and the lowest RMSE possible.</p><p>Because we had several different types of inputs, they needed to be processed in different ways. This led us to choose the Keras Functional API with a FFNN for our first model. For Rain and Temperature we used StandardScalar(), locations were embedded and genetic data was modified through MinMaxScalar(). The Functional API allowed us to add and remove features for experimentation with the model in addition to hypertuning the layers. The final model had 3 dense layers and only used Ag District, State-County and GE data for feature input (Fig 1). However, this model was limited in analyzing relationships with time, or sequences for weather data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/0*Bs0Vqz6Y3s8zCmAa" /><figcaption>Figure 1: Example of Dense Layers for FFNN using Transactional API with comments for experimentation</figcaption></figure><p>In order to assess the sequences for weather, we upgraded to a Transformer model using 2 attention blocks with self-attention. We continued to use the normalized inputs from the FFNN and designed the model to easily add or remove features for experimentation. Tuning included testing additional attention blocks, feature engineering, and neural network layers. Self-attention and dropout were also tested. For this model we used features including year, State-County, GE data, rain and temperature.</p><h3>MHA Transformer Architecture and Technical Details</h3><p>Our transformer model employed a sophisticated dual multi-head attention (MHA) architecture specifically designed to capture temporal relationships in agricultural data. The implementation featured two sequential attention blocks, each with configurable attention heads, followed by 15% dropout, residual connections (Add layers), and layer normalization — as shown in our actual model code (Fig 2). The first attention block processes sequential climate data (temperature and precipitation), while the second attention block operates on the output of the first, creating hierarchical feature learning. Positional encoding enabled the model to understand monthly sequences, distinguishing between January precipitation patterns versus July patterns. The model concatenates county embeddings, static features, and the flattened sequential output before passing through a deep feedforward network (1024 → 512 → 64 units with 50%, 20%, and 30% dropout respectively) for final yield prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lDFVtvnsrWiQf9ED" /><figcaption>Figure 2: Neural Network for our Transformer Model including attention layers, FF layers and hidden layers</figcaption></figure><h3>Results</h3><p>The baseline model, which predicted the historical mean yield, produced a root mean squared error (RMSE) of 39.57 bu/ac. The feedforward neural network (FFNN) reduced this to 20.79 bu/ac by leveraging location and genetically engineered (GE) crop features. The transformer model, which built on the FFNN by incorporating sequential climate data, achieved the best performance with an RMSE of 18.02 bu/ac and a mean absolute error (MAE) of 14.37 bu/ac — representing a 54% improvement over the baseline.</p><p>Our transformer model achieved an RMSE of <strong>18.02 bu/ac</strong>, placing it squarely in the same performance range as some of the best recent image-based corn yield prediction models. For example, <em>Enhancing corn yield prediction: Optimizing data quality or model complexity?</em> (Zhou et al., 2024) used high-resolution RGB imagery and advanced deep learning architectures, reporting RMSEs between <strong>16.54–18.42 bu/ac</strong> across Random Forest, Gradient Boosting, and vision transformer approaches. Notably, our model reached this level of accuracy without relying on expensive or computationally intensive imagery, instead using only climate, location, and genetically engineered crop data — highlighting the power of combining domain-specific tabular features with sequence-aware modeling.</p><p>Models and the full paper can be found here: <a href="https://github.com/aestew/ML_corn_yield_analysis/tree/main">https://github.com/aestew/ML_corn_yield_analysis/tree/main</a></p><p>Team Githubs:</p><p>Chad Adelman: <a href="https://github.com/chadadelman">https://github.com/chadadelman</a><br>Michael White: <a href="https://github.com/mwhite25">https://github.com/mwhite25</a><br>Yoko Morishita: <a href="https://github.com/myoko-codes">https://github.com/myoko-codes</a></p><p>This was a project for our ML class in <a href="https://medium.com/u/255942f15b34">UC Berkeley</a> MIDS.<br>ChatGPT 5.0 was used to summarize and help write this article</p><p><strong>Citation:</strong><br> Zhou, Y., Ma, S., Zhang, H., &amp; Aakur, S. (2024). <em>Enhancing corn yield prediction: Optimizing data quality or model complexity?</em> Agricultural and Forest Meteorology, [Available online 28 November 2024].</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d69190b65464" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>