replicate precision@k score with predict method (rather than predict_rank)

[tinawenzel]

Hi, I tried to replicate the precision@k score resulting from the precision_at_k method using the predict method.

The precision_at_k method is based on predict_rank, but since I have many items to rank for each user, the predict method is more suitable/faster. Clearly, whether one is using predict_rank or predict should not change the precision@k score, but I was unable to replicate the score I get from precision_at_k (based on predict_rank) with the predict method.

In fact the evaluation scores from the predict method are always worse than the evaluation scores derived by the precision_at_k method included in the package. Why is that?

Below is an example using open source data. For simplicity, I'm using only a fraction of the data, a basic model without features, known positives are not removed (train_data argument is not specified in precision_at_k).

Why is this important: The predict method is more suitable in cases where many items need to be ranked, which is my use case as well. Also, I want to calculate ndcg for evaluation and if I can replicate the prec@k score with predict, I know the post-processing of the predictions is correctly set up and I can just change the metric.

    from lightfm import LightFM
    from scipy.sparse import coo_matrix as sp
    import numpy as np
    import pandas as pd
    import matplotlib.pyplot as plt
    import os
    import zipfile
    import csv
    import requests
    import json
    from itertools import islice
    from lightfm.data import Dataset
    from lightfm import LightFM
    from lightfm.evaluation import precision_at_k
    from lightfm.cross_validation import random_train_test_split 

   ######################################
   #                                   
   #  Fetching the training data               
   #                                   
   ######################################

def _download(url: str, dest_path: str):

    req = requests.get(url, stream=True)
    req.raise_for_status()

    with open(dest_path, "wb") as fd:
        for chunk in req.iter_content(chunk_size=2 ** 20):
            fd.write(chunk)

def get_data():

    ratings_url = ("http://www2.informatik.uni-freiburg.de/" "~cziegler/BX/BX-CSV-Dump.zip")

    if not os.path.exists("data"):
        os.makedirs("data")

        _download(ratings_url, "data/data.zip")

    with zipfile.ZipFile("data/data.zip") as archive:
        return (
            csv.DictReader(
                (x.decode("utf-8", "ignore") for x in archive.open("BX-Book-Ratings.csv")),
                delimiter=";",
            ),
            csv.DictReader(
                (x.decode("utf-8", "ignore") for x in archive.open("BX-Books.csv")), delimiter=";"
            ),
            csv.DictReader(
                (x.decode("utf-8", "ignore") for x in archive.open("BX-Users.csv")), delimiter=";"
            ),
        )

def get_ratings():
    return get_data()[0]

def get_book_features():
    return get_data()[1]

def get_user_features():
    return get_data()[2]

# small dataset
udf = pd.DataFrame([x['User-ID'] for x in get_ratings()])
iid = pd.DataFrame([x['ISBN'] for x in get_ratings()])
frames = [udf, iid]

# susample user list
user_set = set([x['User-ID'] for x in get_ratings()])
user_samples = list(user_set)[:800]

train_df = pd.concat(frames, axis=1)
train_df.columns = ['user_id','item_id']
print(train_df.shape)
train_df = train_df[train_df.user_id.isin(user_samples)]
print(train_df.shape)

book_features = [(x['ISBN'], [x['Book-Author']]) for x in get_book_features() if x['ISBN'] in train_df.item_id.unique().tolist()]
user_features = [(x['User-ID'], [x['Age']]) for x in get_user_features() if x['User-ID'] in train_df.user_id.unique().tolist()]

dataset = Dataset()
dataset.fit(train_df.user_id.tolist(),
            train_df.item_id.tolist())
num_users, num_items = dataset.interactions_shape()
print('Num users: {}, num_items {}.'.format(num_users, num_items))

dataset.fit_partial(users=train_df.user_id.tolist(),
                    items=train_df.item_id.tolist(),
                    item_features=[j[0] for i,j in book_features],
                    user_features=[j[0] for i,j in user_features])

#######################
#                                                
#  Building the Model         
#                                               
######################

dataset = Dataset()
dataset.fit(train_df.user_id.unique().tolist(),
            train_df.item_id.unique().tolist())
num_users, num_items = dataset.interactions_shape()
print('Num users: {}, num_items {}.'.format(num_users, num_items))

dataset.fit_partial(users=train_df.user_id.unique().tolist(),
                    items=train_df.item_id.unique().tolist(),
                    item_features=[j[0] for i,j in book_features],
                    user_features=[j[0] for i,j in user_features])

(interactions, weights) = dataset.build_interactions(((i,j) for i,j in zip(train_df.user_id, train_df.item_id)))
print(repr(interactions))

(train, test) = random_train_test_split(interactions=interactions, test_percentage=0.2)

item_features = dataset.build_item_features((book_features))
print(repr(item_features))

user_features1 = dataset.build_user_features((user_features))
print(repr(user_features1))

mapp = dataset.mapping()
dict_user_id = mapp[0]
dict_item_id = mapp[2]

user_list = list(dict_user_id.keys())
items_list = list(dict_item_id.keys())
items =np.array(items_list)

data = {
         'train_cols': items,
         "train": train,
         'test_cols': items,
         "test": test,
         "item_features": item_features, 
         "user_features": user_features1
         }

#############################
#                               
#  Training the Model       
#                               
#############################

model = LightFM(loss='warp')

model.fit(data['train'], 
          #item_features=data['item_features'], 
          #user_features=data['user_features']
         )

### model performnce evaluation
pak = precision_at_k(model,
                      test_interactions = data['test'],
                      #train_interactions = data['train'],
                      #item_features=data['item_features'], 
                      #user_features=data['user_features']
                    ).mean()
print("precision@10 : {}".format(pak))

This gives precision@10 : 0.004322766792029142. Under the hood, the precision@k used the predict_rank method which generates the precision@k like this:

ranks = model.predict_rank(test_interactions=data['test'],
                           #train_interactions=data['train'],
                           #item_features=data['item_features'], 
                           #user_features=data['user_features'],
                           num_threads=32,
                           check_intersections=True)
ranks.data = np.less(ranks.data, 10, ranks.data)
precision = np.squeeze(np.array(ranks.sum(axis=1))) / 10
precision = precision[data['test'].getnnz(axis=1) > 0]
print('prec@10: {}'.format(precision.mean()))

Just to demonstrate that this gives precision@10 : 0.004322766792029142.

############################################
#                               
# Replicate precision using the predict method      
#                               
############################################
mapp = dataset.mapping()
dict_user_id = mapp[0]
dict_item_id = mapp[2]

d_user_pred = {}

for user in dict_user_id.keys():
    d_user_pred[user] = []

for uid, i in dict_user_id.items():   
    known_positives_ids = data['train_cols'][data['train'].tocsr()[i].indices]
    #print('known positives:{}'.format(known_positives_ids))
    scores = model.predict(user_ids = i, 
                           item_ids = np.arange(len(dict_item_id)),
                           #user_features=user_features,
                           #item_features=item_features
                          )

    # get top recommendations
    top_items_ids = data['train_cols'][np.argsort(-scores)]

    # exclude known positives from recommendations
    top_items_ids = np.array(list(set(top_items_ids) - set(known_positives_ids)))
    print('top_items_ids:{}'.format(top_items_ids[:5]))
    d_user_pred[uid] = top_items_ids

##################################
#
# Precision@k evaluation
#
##################################

# get predictions df
df  = pd.DataFrame.from_dict(d_user_pred, orient='index').iloc[:,:20]
df['user_id'] = df.index
df = df.melt(id_vars='user_id')
df.columns = ['user_id','rank','item_id']
pred_df = df.groupby('user_id').aggregate(lambda tdf: tdf.tolist()).reset_index()
pred_df.columns = ['user_id','rank','predictions']

# get ground truth df
t = pd.DataFrame(data['test'].todense(), columns=items_list)
t['user_id'] = user_list
t = t.melt(id_vars='user_id')
t = t[t.value==1].drop('value',axis=1)
t.columns = ['user_id','item_id']
actual_df = t.groupby('user_id').aggregate(lambda tdf: tdf.tolist()).reset_index()
actual_df.columns = ['user_id','actual']

# generate eval_df
eval_df = pred_df.merge(actual_df,on='user_id',how='left')
eval_df = eval_df[eval_df.actual.notnull()]

def precision(actual, predictions, k):
    """ Fraction of retrieved documents @k that are relevant."""
    return len(set(actual) & set(predictions[:k])) / k

eval_df['prec'] = eval_df.apply(lambda row : precision(actual=row['actual'], 
                                                       predictions=row['predictions'], 
                                                       k=10), axis = 1) 
eval_df.prec.mean()

Which gives 0.0005763688760806917.

So in summary, the predict_rank gives precision@k score = 0.004322766792029142 and the predict method gives precision@k score=0.0005763688760806917.

[LeviiBereg]

Hello! I believe the problem is in the order of applied operations in the lines

# get top recommendations
top_items_ids = data['train_cols'][np.argsort(-scores)]

# exclude known positives from recommendations
top_items_ids = np.array(list(set(top_items_ids) - set(known_positives_ids)))

The conversion of top_items_ids into a set set(top_items_ids) completely messes an order of sorted by rank items. I propose using a list comprehension in order to exclude known items known_positives_ids from the top_items_ids like

top_items_ids = [item_id for item_id in top_items_ids if item_id not in known_positives_ids]

[tinawenzel]

Oh, of course. Thanks for pointing that out.

[ppfreitas]

Follow up question if anyone still follows this thread:

Why does measuring performance using the predict method solution given above yields the same results as the predict_rank method without setting the train_interactions argument?

The way I understand it, using the train_interactions argument in the precision_at_k function would be equivalent to the excluding known positives from recommendations step, but I guess it is not.

[ppfreitas]

Follow up question if anyone still follows this thread:

Why does measuring performance using the predict method solution given above yields the same results as the predict_rank method without setting the train_interactions argument?

The way I understand it, using the train_interactions argument in the precision_at_k function would be equivalent to the excluding known positives from recommendations step, but I guess it is not.

Btw, I made the following change to the code and now it work as I expected:

known_positives_ids = data['train_cols'][data['train'].tocsr()[i].indices] known_positives_ids = list(known_positives_ids)

In the original code the list comprehension filter was not working due to known_positives_ids being a pd.Series and not a list.

[esraagamal-111]

is precision@10 : 0.004322766792029142 is good enough for the model , actually I'm working on another model & the best precision about 0.012 so What's your opinion ?