ALS.py

from __future__ import division
import numpy as np
import pandas as pd
import scipy.sparse as sps
import implicit as impl

target_users = np.genfromtxt("target_users.csv", delimiter="\t", dtype=np.dtype(int), skip_header=1)
usersals = pd.read_csv("user_profile.csv", delimiter="\t")
itemsals = pd.read_csv("item_profile.csv", delimiter="\t")
interactions = pd.read_csv("training_data.csv", delimiter="\t")
uds_all = usersals['user_id'].values

def interagiti(user_id):
    return np.unique(interactions[interactions['user_id'] == user_id]['item_id'].values)

attivi = [u for u in uds_all if interagiti(u).size>0]

# PRE-PROCESSING

# drop useless columns
inter = interactions.copy()
inter = inter.drop(['interaction_type', 'created_at'], 1)

# count duplicates
inter = inter.groupby(['user_id', 'item_id']).size().reset_index()
inter = inter.rename(columns={0: 'count'})

# users and items dict : associate user/item id with index
users_dict = dict(zip(usersals['user_id'], usersals.index))
items_dict = dict(zip(itemsals['id'], itemsals.index))
inv_users_dict = {v: k for k, v in users_dict.items()}  # inverse of users_dict
inv_items_dict = {v: k for k, v in items_dict.items()}  # inverse of items_dict

# add users/items indexes columns
inter['user_index'] = [users_dict.get(i) for i in inter['user_id']]
inter['item_index'] = [items_dict.get(i) for i in inter['item_id']]

# non-active items ids/indexes
items_nact_ids = itemsals[itemsals['active_during_test'] == 0]['id'].drop_duplicates().values
items_nact = [items_dict.get(i) for i in items_nact_ids]


# URM : user-rating matrix
urm = sps.csr_matrix((inter['count'], (inter['user_index'], inter['item_index'])))


#ALS: Alternative Least Squares

alpha = 40
factors = 300
regularization = 0.01
iterations = 20

matr = sps.lil_matrix((len(usersals), urm.shape[1]))

def save_sparse_csr(filename,array):
    np.savez(filename,data = array.data ,indices=array.indices,
             indptr =array.indptr, shape=array.shape )


user_vecs, item_vecs = impl.alternating_least_squares((urm*alpha).astype('double'), factors, regularization, iterations)

l = len(uds_all)

for u_id in uds_all:
    u = users_dict.get(u_id)
    rec = user_vecs[u, :].dot(item_vecs.T)
    rec[items_nact] = 0
    cols = np.argsort(rec)[::-1][:1000]
    rec = rec[cols]
    rec = np.true_divide(rec,np.max(rec))
    matr[u, cols] = rec
    l -= 1
    print(l)

save_sparse_csr('ALS_ALL2', matr.tocsr())