Data Drift in ML Production. How to measure? One way is Kullback-Leibler (KL) divergence 👇 🧠 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞? KL divergence is a measure that quantifies how one probability distribution differs from another reference distribution. 🤔 𝐇𝐨𝐰 𝐢𝐬 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐔𝐬𝐞𝐝 𝐢𝐧 𝐃𝐚𝐭𝐚 𝐃𝐫𝐢𝐟𝐭 𝐌𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠? When an ML model is deployed, KL divergence can continuously track the difference between the training data distribution (reference) and the live (production data). ⚠️ If KL divergence increases beyond a certain threshold, it signals that the data has drifted, which could mean that the model's performance is at risk. 🟢 𝐀𝐝𝐯𝐚𝐧𝐭𝐚𝐠𝐞 𝐨𝐟 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞: KL divergence is highly sensitive to even minor shifts in distribution which may be useful for data change sensitive applications. It makes it handy for detecting subtle forms of data drift early if required. 🔴 𝐃𝐢𝐬𝐚𝐝𝐯𝐚𝐧𝐭𝐚𝐠𝐞 𝐨𝐟 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞: KL divergence may flag small deviations (e.g., outliers or noise) as data drift, leading to false alarms. ♻️ Share it, so your network knows what you are interested in! P.S. Do you know other ML use cases for KL divergence?by Amin on TwosData Drift in ML Production. How to measure? One way is Kullback-Leibler (KL) divergence 👇 🧠 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞? KL divergence is a measure that quantifies how one probability distribution differs from another reference distribution. 🤔 𝐇𝐨𝐰 𝐢𝐬 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐔𝐬𝐞𝐝 𝐢𝐧 𝐃𝐚𝐭𝐚 𝐃𝐫𝐢𝐟𝐭 𝐌𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠? When an ML model is deployed, KL divergence can continuously track the difference between the training data distribution (reference) and the live (production data). ⚠️ If KL divergence increases beyond a certain threshold, it signals that the data has drifted, which could mean that the model's performance is at risk. 🟢 𝐀𝐝𝐯𝐚𝐧𝐭𝐚𝐠𝐞 𝐨𝐟 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞: KL divergence is highly sensitive to even minor shifts in distribution which may be useful for data change sensitive applications. It makes it handy for detecting subtle forms of data drift early if required. 🔴 𝐃𝐢𝐬𝐚𝐝𝐯𝐚𝐧𝐭𝐚𝐠𝐞 𝐨𝐟 𝐊𝐋 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞: KL divergence may flag small deviations (e.g., outliers or noise) as data drift, leading to false alarms. ♻️ Share it, so your network knows what you are interested in! P.S. Do you know other ML use cases for KL divergence?