Machine Learning Q&A: Data Imputation, Feature Selection, Recommendation Systems and More

Q1: How can missing user behavior data be inferred from relatively complete data?

If the missing behavior data are numeric, build a predictive model using the existing data, train it, and fill the gaps with predictions. For non‑numeric actions (e.g., event logs), analyze the distribution of existing data and fill missing parts with random functions plus rule‑based logic.

Q2: How to perform feature extraction and remove useless features? Which recommendation algorithms does Meituan‑Dianping use?

1. Feature extraction requires understanding business data and logic; transform raw features when needed. Feature selection can use tree models, L1 regularization, etc., with tools like XGBoost. 2. Meituan‑Dianping’s recommendation system uses a recall stage that combines collaborative filtering, location, search queries, real‑time user behavior, etc., and a ranking stage based on Learning‑to‑Rank techniques.

Q3: What is the optimal learning path for beginners in machine learning? Any recommended book list?

Start with Li Hang’s “Statistical Learning Methods” for fundamentals, complemented by Coursera’s “Machine Learning” video course. For video learners, consider Lin Xuantian’s two video series from National Taiwan University. Then practice with simple competition problems (e.g., basic click‑through rate prediction) using tools such as pandas and scikit‑learn.

Q4: How is feature selection usually performed for machine learning models?

Two main categories: (1) Use fixed evaluation metrics like information gain from an ID3 decision tree to rank features; selected features are then fed into the model. (2) Use model feedback: iteratively add features that most improve prediction performance until gains plateau, or iteratively remove features. The first method is fast but lacks feedback; the second yields better results but is slower.

Q5: What is the mathematical principle behind L1 regularization yielding sparse solutions?

From the perspective of gradient descent, the gradient of the L1 norm is constant, so the penalty is independent of the parameter’s location, making it easier for the optimizer to drive some coefficients to zero.

Q6: Are there text classification algorithms that significantly outperform the traditional tf‑idf/word2vec + linear‑SVM/Bayes pipeline?

Strongly recommend TextCNN; if the training data for embeddings are insufficient, word2vec can be used as a substitute, still delivering good performance.

Q7: How should the training sample set for search ranking be selected? Does using only exposed items bias the model?

Training typically uses exposed items, but feature extraction is generalized enough that unexposed results can still be learned effectively. The recall layer aims to reduce the computational load of the ranking stage while preserving user‑relevant results.

Q8: In wide & deep recommendation models, are training labels manually annotated or generated automatically?

Labels for wide & deep models are automatically generated from user behavior data.

Q9: Previously, video target detection used the VIBE algorithm. Are deep‑learning methods now dominant?

Deep‑learning methods now dominate video object detection; the top five entries of the ILSVRC 2016 VID competition all used deep learning. VIBE offers good real‑time performance but lags behind deep‑learning approaches in accuracy.

Q10: Has anyone used Factorization Machines (FM) for click‑through rate prediction? Any open‑source engineering code available?

Various CTR competitions on Kaggle (e.g., Criteo Display Advertising Challenge, Avazu CTR Prediction) provide public code that can be referenced for FM implementations.