Similarity movies in recommendation movies

Course Q&A Machine Learning Specialization Unsupervised Learning, Recommenders, Reinforcement
,
1

In content based filtering, using two nuerol networks to predict user’s movie preference is a cool example.

I have some technical quesitons in two networks, in the end you have V_u and V_m vectors and take inner product.

However, when Ng recommends finding similar movies, he recommends ||V_u - V_m||^2 as a metrics to find similar movies. I think that both are legitimate metrics, but I feel that the inner products to find similarity is more appropriate.

Can you comment on why we don’t use inner product when we want similar movies?
||V_m^k - V_m^i||**2 vs np.dot(V_m^k , V_m^i).

Thanks,

1 Like
2

Hello @yildirimga,

image
image1616×333 38.6 KB

The distance metric is only proportional to the dot product metric only when the norms of the vectors are constant. They are constant if the model normalizes them. Conversely speaking, if they are not normalized, it’s better to stick to the metric used by the training process.

Cheers,
Raymond

1 Like