ikatsov

from gensim.models.doc2vec import TaggedDocument

EMBEDDING_DIM = 200    # dimensionality of user representation

class TaggedDocumentIterator(object):
    def __iter__(self):
        for row in self.df.itertuples():
            yield TaggedDocument(
                     words=dict(row._asdict())['all_orders'].split(),
                     tags=[dict(row._asdict())['user_id']])

ikatsov

orders.sort_values(by=['user_id','order_number','add_to_cart_order'], 
                   inplace=True)

orders_by_uid = orders.groupby("user_id")
                      .apply(lambda order: ' '.join(order['product_id'].tolist()))

orders_by_uid = pd.DataFrame(orders_by_uid,  
                             columns=['all_orders'])
orders_by_uid.reset_index(inplace=True)
orders_by_uid.user_id = orders_by_uid.user_id.astype(str)

ikatsov

from sklearn.preprocessing import MinMaxScaler
mm_scale = MinMaxScaler()

# feature_df is the dataframe with customer features
feature_df_scale = pd.DataFrame(mm_scale.fit_transform(feature_df),
                                columns=feature_df.columns,
                                index=feature_df.index.values)

tsne_doc_features = TSNE(n_components=2, verbose=1, perplexity=30, n_iter=500)
tsne_features_doc = tsne_doc_features.fit_transform(feature_df_scale.values)

ikatsov

from sklearn.metrics import silhouette_samples, silhouette_score

for space_name, space in {'t-SNE': tsne_results, 
                          'original': model.wv.vectors}.items():
    for entity in ['department', 'aisle']:
        s = silhouette_score(space, df_semantic_item[entity], metric="cosine")
        print(f"Score on {space_name} space for {entity}s is {s:.4}")

--------------

ikatsov

from sklearn.manifold import TSNE
from sklearn.metrics import pairwise_distances

# prepare inputs for t-SNE
word_vectors = model.wv
vocab = list(model.wv.vocab.keys())
item2vector_dict = {arg:model.wv[arg] for arg in vocab}
X = pd.DataFrame(item2vector_dict).T.values

# perform t-SNE

ikatsov

def to_product_name(id, columns='product_name'):
    Return products_csv[products_csv.product_id==id][columns].values.tolist()[0]

def most_similar_readable(model, product_id, topn=10):
    similar_list = [(product_id, 1.0)] + model.wv.most_similar(str(product_id), 
                                                               topn=topn)
    return pd.DataFrame([( to_product_name(int(id)), int(id), similarity ) for 
                         (id, similarity) in similar_list],
                        columns=['product', 'product_id', 'similarity'])

ikatsov

from gensim.models import Word2Vec
import multiprocessing as mp

WORD_DIM = 200   # dimensionality of the embedding space
model = Word2Vec(product_corpus, 
                 window=5, 
                 size=WORD_DIM, 
                 workers=mp.cpu_count() - 2, 
                 min_count=100)

ikatsov

order_ds = orders_csv.merge(order_products_csv, 
                            left_on='order_id',
                            right_index=True)

# Creating sequences based on transactions
order_product_list = order_ds.sort_values(
                         ['user_id','order_id','add_to_cart_order'])
                         [['order_id','product_id']].values.tolist()

# Each entry of a corpus is one order represented by 

ikatsov

products.csv
  | product_id | product_name                | aisle_id | department_id |
  | 1          | Chocolate Sandwich Cookies  | 61       | 19            |
  | 2          | All-Seasons Salt            | 104      | 13            |  
  | 3          | Robust Golden Oolong Tea    | 94       | 7             |  
  ...

departments.csv    # coarse categorization
  | department_id  | department |
  | 1              | frozen     |

ikatsov

config = ddpg.DEFAULT_CONFIG.copy()
config["actor_hiddens"] = [512, 512] 
config["critic_hiddens"] = [512, 512]
config["gamma"] = 0.95
config["timesteps_per_iteration"] = 1000
config["target_network_update_freq"] = 5
config["buffer_size"] = 10000
    
trainer = ddpg.DDPGTrainer(config=config, env=SimpleSupplyChain)
for i in range(n_iterations):