14. Under the hood…
Data: post titles, comment text for 30K posts.
Model: word2vec model trained on comment text.
15. Under the hood…
Data: post titles, comment text for 30K posts.
Model: word2vec model trained on comment text.
python
amazon
app
web
vocabulary
apple
Word2Vec space W
web
python
amazon
julia
app
apple
julia
16. Recommendation, Step 1
Recommending posts
1. Map each user interest to a vector.
2. Map each post to a vector.
3. For each interest, recommend the “nearest”
30 posts.
17. Recommendation, Step 1
Recommending posts
1. Map each user interest to a vector.
2. Map each post to a vector.
3. For each interest, recommend the “nearest”
30 posts.
22. Recommendation, Step 1
Recommending posts
1. Map each user interest to a vector.
2. Map each post to a vector.
3. For each interest, recommend the “nearest”
30 posts.
26. Recommendation, Step 1
Recommending posts
1. Map each user interest to a vector.
2. Map each post to a vector.
3. For each interest, recommend the “nearest”
30 posts.
29. Recommendation, Step 1
Recommending posts
1. Map each user interest to a vector.
2. Map each post to a vector.
3. For each interest, recommend the “nearest”
30 posts.
35. Why this setup?
• Initial thought: some kind of topic modeling.
• User input is inflexible (topics are incoherent if we
use more than 10).
• Ideal use case is streaming, but number of topics
must be chosen manually, and “meanings” must
be assigned to topics manually.
• Similar remarks apply to clustering + dimension
reduction.