Yingjue Zhou

Forecasting what matters: TimesFM-2.5 on the VN2 Inventory Challenge (with covariates) Short horizons. Intermittent demand. Real retail constraints. I just ran Google Research’s TimesFM 2.5 across all 599 SKUs of the VN2 Inventory Planning challenge—and then layered a small, transparent covariates step (store, product hierarchy, calendar, drift index) on top. What I learned: - Strong zero-/few-shot prior: TimesFM 2.5 is already a solid baseline without any training. - Covariates help where planners feel it most: a simple xreg layer reduced error and pulled bias closer to zero—crucial for inventory decisions. - Baselines still matter: a seasonal MA remains a great guardrail for calibration; routed SES/Croston/TBATS helps on very intermittent SKUs; LightGBM is competitive when you engineer horizon-safe features. - Keep it legible: the whole stack is understandable enough to discuss with planners (and to debug when reality shifts). Why post this? Because foundation models for time series are becoming practical. Kudos to Google Research (Yossi Matias) for pushing the field with TimesFM, and to Nicolas Vandeput for making datasets like VN2 (and earlier VN1) accessible for honest benchmarking. If you’re exploring retail/supply-chain forecasting: start with a strong prior, add the covariates you trust, and make bias everyone’s business. Happy to share notebooks and code. — Resources (in comments): repo + notebook, VN2 challenge, and a pointer to VN1. #TimeSeries #Forecasting #GoogleResearch #TimesFM #RetailAnalytics #InventoryPlanning #FoundationModels #DataScience #VN2 #VN1 #OpenSource

42

4 Comments

In the fast-paced e-commerce landscape, category popularity can make or break a platform. Popular categories thrive with more interactions and room for learning, while less-frequented ones often struggle. In my latest video, I dissected KDD 2021's paper by Meituan, a leading Chinese e-commerce platform, titled 'A Dual Augmented Two-Tower Model For Online Large-scale Recommendation.' This research explores the significance of category popularity in e-commerce success. It introduces an innovative approach to boost underdog categories using a Category-Alignment Loss mechanism, enhancing item embeddings. Beyond category alignment, the paper also addresses bridging the gap between towers in the two-tower model, creating a more comprehensive recommendation system. This research has real-world applications, particularly in Reddit's Ads Recommendation Systems. Find out more details about it in my YT channel where I dived into these innovative approaches. Video Link: youtu.be/UhpbTSbi3lI Channel Link: youtube.com/@datatrek #datatrek #datascience #machinelearning #statistics #deeplearning #ai

1

When building an AI chatbot to retrieve order history in an e-commerce setting, consistency matters. This is why we set the model temperature low — to reduce randomness and deliver reliable, repeatable answers. LLMs are inherently probabilistic. Temperature controls how likely the model is to choose less probable outputs. Lower temperature implies more determinism; higher temperature allows for more creativity. It’s a key parameter when designing AI agents for enterprise applications. For example, consider this order status query: User: “What’s the status of my last order?” Response 1: “Your last order for a wireless mouse was delivered on July 28.” Response 2: “Looks like your recent order arrived a few days ago. Let us know if you need anything else!” Both are correct — but in enterprise workflows, variation like this can cause confusion. That’s why tuning model temperature is critical. A few seats are still available in Building AI Agents in Enterprise course that Ajay Shenoy and Abhijith Neerkaje are teaching. Link to register is in the comments below.

69

1 Comment

New short course: DSPy: Build and Optimize Agentic Apps DSPy is a powerful open-source framework for automatically tuning prompts for GenAI applications. In this course, you'll learn to use DSPy, together with MLflow. This is built in partnership with Databricks and taught by Chen Qian, co-lead of the DSPy framework. Many AI builders spend hours hand-tuning prompts. When given a set of evals, DSPy automates this process. It’s especially useful for optimizing prompts, including few-shot prompts, in complex agentic AI workflows. Further, if you switch an application to a newer LLM, performance can degrade if your prompts were optimized to the previous model. DSPy automatically optimizes the entire system for the new LLM as well, using just a few evaluation examples. This course teaches DSPy works, and best practices for using it. You’ll write programs using DSPy’s signature-based programming model, debug them with MLflow tracing -- to gain visibility into how different parts of a pipeline, as well as how the overall system, are performing -- and automatically improve their accuracy with DSPy Optimizer. Please sign up here: https://lnkd.in/gdjae8AX

4,671

190 Comments

Inference at scale is evolving. The future is composable systems where offline batch, online batch (near-realtime), and realtime traffic can all be served from a single, unified inference endpoint for most use-cases. This simplifies your architecture, reduces operational overhead, and increases throughput. Our new tutorial series will show you how to build it step by sep. In the first installment, Erik Saarenvirta walks you through creating a scalable image classification system on GKE that can be adapted to a variety of use-cases. Leave us feedback in comments or ask questions we are happy to answer. Kent Hua Ishmeet Mehta Erik Saarenvirta https://lnkd.in/dVNiUhwE

16

“Good luck with that,” Lei Tang said, when I asked about how data teams can get their data ready for AI. Lei has spent his career leading data science at Lyft, Clari, Walmart Labs - and now as co-founder and CTO of Fabi.ai. He’s talked to hundreds of data leaders. And the one thing they all agree on? The data is never clean. Not at early-stage startups. Not at public companies. Not even in teams drowning in headcount and budget. The myth of “data readiness” - that you have to fix your pipelines and documents and semantics before you do AI - is holding companies back. In this episode, he shares: - Why waiting for perfect data is the wrong strategy for adopting AI - What changes when you assume the data will always be messy - How AI can learn your org’s context from dashboards, behavior, and usage patterns - not from documentation And if you’ve ever felt like you’re stuck in cleanup mode instead of impact mode, give this episode a watch. Full interview in comment below.

36

2 Comments

It is hard to understand Fabric's long-term strategy. On one hand, it was smart to build it around PBI, the leading BI platform. On the other hand, the key to Power BI's success in dethroning Tableau was its' extremely affordable licensing. Competitive licensing is not an element that Fabric's always-on pricing model can boast of at current rates, especially when the likes of Databricks, Snowflake, and other rising platforms are around. Perhaps Microsoft will be content with Fabric always being the 3rd/4th best option?

39

32 Comments

Happy Monday Morning with SVM & ☕ Biggest roadblock in ML applications ? Cost and ROI trade off! Let's kick off with Support Vector Machines - a classic ML algorithm that still packs a serious punch in 2025 Why SVMs remain popular: 1. Kernel Flexibility - SVMs shine with the kernel trick, transforming non-linearly separable data into higher dimensions where a clean separation exists. The RBF kernel (as shown in our Bayesian Optimization example) handles complex decision boundaries without explicitly computing high-dimensional transformations. 2. Robust with Small Datasets - Unlike deep learning models that hunger for massive datasets, SVMs deliver strong performance with limited samples. Their margin maximization approach finds the optimal separating hyperplane even when data is scarce. Real-world impact: 🏥 WHO (World Health Organization) - Leverages SVMs for disease classification and outbreak prediction, where medical datasets are often limited but accuracy is critical. 🌍 NASA - Uses SVMs for satellite image classification and anomaly detection in space mission data, where the cost of misclassification is extraordinarily high. Keep these nuances in mind: ⚠️ Hyperparameter sensitivity - C (regularization) and gamma (kernel coefficient) dramatically affect performance. As our Bayesian Optimization code shows, finding optimal values requires sophisticated search strategies - not just trial and error. ⚠️ Computational cost - Training time scales poorly beyond 10K samples. For large datasets, consider SGD-based alternatives or ensemble methods that parallelize better. Cost & Environmental Impact: While SVMs are computationally lighter than modern deep learning, hyperparameter tuning through grid search generates significant compute waste. Bayesian Optimization (featured in our example) reduces this by ~10x, cutting both cloud costs and carbon footprint. Pro tip: Start with C=1.0 and gamma='scale' as sensible defaults, then optimize only if needed #MachineLearning #DataScience #SVM #AI #MondayMotivation Credits: Image by OpenAI | Content curated by GrokAI Pardon the typos in the image :)

4