lpraat / recsys Goto Github PK

Whenever we add together multiple similarity matrix(of the same size) like we do in ItemKNN I think it's better to apply KNN on the resulting sum rather than applying it singularly on each similarity matrix.

Why I propose this:

• using this approach we will have just one "knn" parameter to tune, that is the one applied last. This could help even in the future when maybe we will put and sum together other similarity matrix from other kind of algorithms(like SLIM similarity). Moreover, even right now we tune just the item knn and we don't bother about the artist and album knn
• in the current code aren't we increasing the density of the final similarity matrix?

I think this modification is worth trying and we can see if it actually improves or degrades performance.

I would change, for example, this

 s += cosine_similarity(
                    feature,
                    alpha=feature_alpha,
                    asym=feature_asym,
                    h=feature_h,
                    knn=feature_knn,
                    qfunc=feature_qfunc,
                    dtype=np.float32
                ) * feature_w

to

s+= cosine_similarity(
                    feature,
                    alpha=feature_alpha,
                    asym=feature_asym,
                    h=feature_h,
                    # knn here is removed
                    qfunc=feature_qfunc,
                    dtype=np.float32
                ) * feature_w

# Here after computing the sum we apply knn
apply_knn(s)

Speedup Slim BPR using cython both in the sampling and in the training