The document details a news recommendation system utilizing machine learning (ML) and natural language processing (NLP) to understand user interests and provide tailored news content. It covers the architecture, training processes, and various NLP techniques including text processing, clustering, and topic modeling. The document also highlights the application's potential for personalized recommendations and ongoing advancements in the field.

1. No more bad news!
News recommendation with ML and NLP.
Samia Khalid and Simon Lia-Jonassen
NTNU Cogito
March 7th, 2019

2. Contents 00 Introduction
01 Recommender architecture
02 Natural language processing
03 Recommendation model training
04 Demo and further work

3. Understand Content of news I read.
Learn my Interests over time.
Recommend news that interest me.
Introduction to News Recommendation

4. Implements three parts:
• Frontend and backend controllers.
• Feed provider and logging.
• NLP, ML and exploration workflow.
News recommender in a nutshell
https://github.com/s-j/goodnews

5. Content and feedback signals

6. Natural
Language
Processing

7. Natural Language Processing

8. 1. Text Processing
2. Clustering
3. Topic Extraction
Natural Language Processing and Exploration

9. 1. Text Processing using spaCy
Leading open-source library for advanced NLP

10. a. Tokenization
b. Part Of Speech Tagging
c. Lemmatization
d. Stop words
1. Text Processing using spaCy

11. 1. Recognizes a sentence and
assigns a syntactic structure to
it
• “Who is the AI research director?”
2. spaCy provides a built-in
visualizer
1. Text Processing using SpaCy
Dependency Parsing

12. 1. Locate and classify named entities in text into pre-defined categories
2. Can help to answer questions like:
• “Which people, companies and products is the user interested in?”
1. Text Processing using SpaCy
Entity Recognition

13. 1. Part-of-speech Tagging:
• assigns parts of speech to each token
such as noun, verb, adjective, etc.
2. spaCy uses a statistical model to
make a prediction of which tag or
label most likely applies in the
given context
1. Text Processing using SpaCy
Distribution of POS Tags

14. 1. Text Processing using SpaCy
Word Probabilities: finding the most improbable words (noisy data)

15. 1. Text Processing using SpaCy
Analyzing top unigrams in clicked articles vs all articles (considering only PROPN and NOUNS)

16. 1. Word Vectors as input:
• 300 dimensional vectors to represent words in
numerical form
2. K-Means needs the number of cluster as
parameter:
• Try out different values until satisfied
• Can use silhouette score and distortion as metric
3. PCA for visualizing the results in 2-D
2. K-Means Clustering

17. 2. K-Means Clustering
Note
Clusters for the full set of articles

18. 1. LDA considers two things:
• Each document in a corpus is a weighted combination of several topics, e.g.,
doc1-> 0.1 finance + 0.2 science + 0.5 * technology,…
• Each topic has its collection of representative keywords, e.g.,
technology -> [‘computer’, ‘microsoft’, ‘google', ...]
3. Topic Modeling: LDA

19. 2. The two probability distributions that the algorithm tries to approximate,
starting from a random initialization until convergence:
• For a given document, what is the distribution of topics that describe it?
• For a given topic, what is the distribution of its words or what is the importance (probability) of
each word in defining the topic nature?
3. Topic Modeling: LDA

20. 3. Topic Modeling: LDA
Interactive Topics
Visualization with
pyLDAvis

21. Recommendation
Model
Training

22. 1. Join requests and feedback logs.
• Alternative: use a third-party dataset.
2. Use #clicks > 0 as a positive label.
• Alternative 1: use #clicks / #views
• Alternative 2: use click order
• Alternative 3: get explicit feedback
Model training
Preprocessing

23. 1. Use title and description to get:
• A bag of named entities such as person or org
(using spaCy).
• A bag of key terms from the semantic network
(using Textacy).
• A normalized sum over key term embedding
vectors found in GoogleNews word2vec dataset.
2. Hold out 20% of items for testing.
Model training
NLP features and Train/Test split

24. 1. One-hot-encode entities to get a sparse vector.
2. Compensate popularity skew using
Inverse Document Frequency (IDF).
3. Train a classifier using Gradient Boosting
Decision Trees (GBDT).
Note that we have a small and very skewed,
noisy dataset so we are not expecting good
classification performance.
Model training
Pipeline based on entities

25. 1. Hash-merge features into 100 buckets.
2. Train a GBDT classifier.
Model training
Pipeline based on semantic key terms

26. Just use logistic regression right away.
• This gives us a more relaxed prediction with a
much higher number of true positives but
also false negatives.
Model training
Pipeline based on embedding vectors

27. Model training
Stacking and beyond
It is possible to combine features and models...

28. 1. Get NLP features for a ranking candidate.
• Equivalent to the preprocessing step in training.
2. Get "click probability" from the loaded pipeline
and use this value for ranking.
Model application
Using a trained model

29. Demo time!

30. • More data and NLP/ML advancements
• Personalized recommendation
• Incremental and online learning
• Social signals and behaviors
Further work

31. © Copyright Microsoft Corporation. All rights reserved.
Thank you!