Skip to content
Extraits de code Groupes Projets
Valider 1224faf3 rédigé par Adrien Payen's avatar Adrien Payen
Parcourir les fichiers

update evaluator

parent 2c0d6a12
Aucune branche associée trouvée
Aucune étiquette associée trouvée
Aucune requête de fusion associée trouvée
Masquer les modifications d'espaces
En ligne Côte à côte
Affichage de
avec 35 ajouts et 17 suppressions
evaluator.ipynb
+ 35
− 17
Voir le fichier @ 1224faf3
%% Cell type:markdown id:a665885b tags:
# Evaluator Module
The Evaluator module creates evaluation reports.
Reports contain evaluation metrics depending on models specified in the evaluation config.
%% Cell type:code id:6aaf9140 tags:
``` python
# reloads modules automatically before entering the execution of code
%load_ext autoreload
%autoreload 2
# imports
import numpy as np
import pandas as pd
# local imports
from configs import EvalConfig
from constants import Constant as C
from loaders import export_evaluation_report
from loaders import load_ratings
# New imports
from surprise.model_selection import train_test_split
from surprise import accuracy
from surprise.model_selection import LeaveOneOut
from collections import Counter
```
%% Cell type:markdown id:d47c24a4 tags:
# 1. Model validation functions
Validation functions are a way to perform crossvalidation on recommender system models.
%% Cell type:code id:d6d82188 tags:
``` python
# -- implement the function generate_split_predictions --
def generate_split_predictions(algo, ratings_dataset, eval_config):
"""Generate predictions on a random test set specified in eval_config"""
# Spliting the data into train and test sets
trainset, testset = train_test_split(ratings_dataset, test_size=eval_config.test_size)
# Training the algorithm on the train data set
algo.fit(trainset)
# Predict ratings for the testset
predictions = algo.test(testset)
return predictions
# -- implement the function generate_loo_top_n --
def generate_loo_top_n(algo, ratings_dataset, eval_config):
"""Generate top-n recommendations for each user on a random Leave-one-out split (LOO)"""
# Create a LeaveOneOut split
loo = LeaveOneOut(n_splits=1)
for trainset, testset in loo.split(ratings_dataset):
algo.fit(trainset) # Train the algorithm on the training set
anti_testset = trainset.build_anti_testset() # Build the anti test-set
predictions = algo.test(anti_testset) # Get predictions on the anti test-set
top_n = {}
for uid, iid, _, est, _ in predictions:
if uid not in top_n:
top_n[uid] = []
top_n[uid].append((iid, est))
for uid, user_ratings in top_n.items():
user_ratings.sort(key=lambda x: x[1], reverse=True)
top_n[uid] = user_ratings[:eval_config.top_n_value] # Get top-N recommendations
anti_testset_top_n = top_n
return anti_testset_top_n, testset
def generate_full_top_n(algo, ratings_dataset, eval_config):
"""Generate top-n recommendations for each user with full training set (LOO)"""
full_trainset = ratings_dataset.build_full_trainset() # Build the full training set
algo.fit(full_trainset) # Train the algorithm on the full training set
anti_testset = full_trainset.build_anti_testset() # Build the anti test-set
predictions = algo.test(anti_testset) # Get predictions on the anti test-set
top_n = {}
for uid, iid, _, est, _ in predictions:
if uid not in top_n:
top_n[uid] = []
top_n[uid].append((iid, est))
for uid, user_ratings in top_n.items():
user_ratings.sort(key=lambda x: x[1], reverse=True)
top_n[uid] = user_ratings[:eval_config.top_n_value] # Get top-N recommendations
anti_testset_top_n = top_n
return anti_testset_top_n
def precomputed_information(movie_data):
""" Returns a dictionary that precomputes relevant information for evaluating in full mode
Dictionary keys:
- precomputed_dict["item_to_rank"] : contains a dictionary mapping movie ids to rankings
- (-- for your project, add other relevant information here -- )
"""
# Initialize an empty dictionary to store item_id to rank mapping
item_to_rank = {}
# Calculate popularity rank for each movie
ratings_count = movie_data.groupby('movieId').size().sort_values(ascending=False)
# Assign ranks to movies based on their popularity
for rank, (movie_id, _) in enumerate(ratings_count.items(), start=1):
item_to_rank[movie_id] = rank
# Create the precomputed dictionary
precomputed_dict = {}
precomputed_dict["item_to_rank"] = item_to_rank
return precomputed_dict
def create_evaluation_report(eval_config, sp_ratings, precomputed_dict, available_metrics):
""" Create a DataFrame evaluating various models on metrics specified in an evaluation config.
"""
evaluation_dict = {}
for model_name, model, arguments in eval_config.models:
print(f'Handling model {model_name}')
algo = model(**arguments)
evaluation_dict[model_name] = {}
# Type 1 : split evaluations
if len(eval_config.split_metrics) > 0:
print('Training split predictions')
predictions = generate_split_predictions(algo, sp_ratings, eval_config)
for metric in eval_config.split_metrics:
print(f'- computing metric {metric}')
assert metric in available_metrics['split']
evaluation_function, parameters = available_metrics["split"][metric]
evaluation_dict[model_name][metric] = evaluation_function(predictions, **parameters)
# Type 2 : loo evaluations
if len(eval_config.loo_metrics) > 0:
print('Training loo predictions')
anti_testset_top_n, testset = generate_loo_top_n(algo, sp_ratings, eval_config)
for metric in eval_config.loo_metrics:
assert metric in available_metrics['loo']
evaluation_function, parameters = available_metrics["loo"][metric]
evaluation_dict[model_name][metric] = evaluation_function(anti_testset_top_n, testset, **parameters)
# Type 3 : full evaluations
if len(eval_config.full_metrics) > 0:
print('Training full predictions')
anti_testset_top_n = generate_full_top_n(algo, sp_ratings, eval_config)
for metric in eval_config.full_metrics:
assert metric in available_metrics['full']
evaluation_function, parameters = available_metrics["full"][metric]
evaluation_dict[model_name][metric] = evaluation_function(
anti_testset_top_n,
**precomputed_dict,
**parameters
)
return pd.DataFrame.from_dict(evaluation_dict).T
```
%% Cell type:markdown id:f7e83d1d tags:
# 2. Evaluation metrics
Implement evaluation metrics for either rating predictions (split metrics) or for top-n recommendations (loo metric, full metric)
%% Cell type:code id:f1849e55 tags:
``` python
# -- implement the function get_hit_rate --
def get_hit_rate(anti_testset_top_n, testset):
"""Compute the average hit over the users (loo metric)
A hit (1) happens when the movie in the testset has been picked by the top-n recommender
A fail (0) happens when the movie in the testset has not been picked by the top-n recommender
"""
hits = 0
total_users = len(testset)
for uid, true_iid, _ in testset:
if uid in anti_testset_top_n and true_iid in {iid for iid, _ in anti_testset_top_n[uid]}:
hits += 1
hit_rate = hits / total_users
return hit_rate
# -- implement the function get_novelty --
def get_novelty(anti_testset_top_n, item_to_rank):
"""Compute the average novelty of the top-n recommendation over the users (full metric)
The novelty is defined as the average ranking of the movies recommended
"""
total_rank_sum = 0
total_recommendations = 0
for uid, recommendations in anti_testset_top_n.items():
for iid, _ in recommendations:
if iid in item_to_rank:
total_rank_sum += item_to_rank[iid]
total_recommendations += 1
if total_recommendations == 0:
return 0 # Avoid division by zero
average_rank_sum = total_rank_sum / total_recommendations
return average_rank_sum
```
%% Cell type:markdown id:1a9855b3 tags:
# 3. Evaluation workflow
Load data, evaluate models and save the experimental outcomes
%% Cell type:code id:704f4d2a tags:
``` python
AVAILABLE_METRICS = {
"split": {
"mae": (accuracy.mae, {'verbose': False}),
"rmse": (accuracy.rmse, {'verbose': False})
},
"loo": {
"hit_rate": (get_hit_rate, {}),
},
"full": {
"novelty": (get_novelty, {}),
}
}
sp_ratings = load_ratings(surprise_format=True)
precomputed_dict = precomputed_information(pd.read_csv("data/tiny/evidence/ratings.csv"))
evaluation_report = create_evaluation_report(EvalConfig, sp_ratings, precomputed_dict, AVAILABLE_METRICS)
export_evaluation_report(evaluation_report)
```
%% Output
Handling model baseline_1
Training split predictions
- computing metric mae
- computing metric rmse
Training loo predictions
Training full predictions
Handling model baseline_2
Training split predictions
- computing metric mae
- computing metric rmse
Training loo predictions
Training full predictions
Handling model baseline_3
Training split predictions
- computing metric mae
- computing metric rmse
Training loo predictions
Training full predictions
Handling model baseline_4
Training split predictions
- computing metric mae
- computing metric rmse
Training loo predictions
Training full predictions
The data has been exported to the evaluation report
mae rmse hit_rate novelty
baseline_1 1.563822 1.787365 0.046729 99.405607
baseline_2 1.535869 1.866364 0.018692 429.942991
baseline_3 0.871233 1.081468 0.037383 99.405607
baseline_4 0.729477 0.926489 0.158879 60.583178
baseline_1 1.517749 1.745787 0.056075 99.405607
baseline_2 1.472806 1.805674 0.000000 429.942991
baseline_3 0.868666 1.076227 0.093458 99.405607
baseline_4 0.713063 0.912046 0.074766 60.349533
%% Cell type:markdown id:9fbf23fd tags:
Analyzing the provided data on different baselines, several observations can be made across various metrics.
Firstly, looking at the Mean Absolute Error (MAE), baseline_4 stands out with the lowest value of 0.713063, indicating superior accuracy in predictions compared to the other baselines. Following closely behind is baseline_3 with a MAE of 0.868666, showcasing commendable precision in its predictions.
Next, considering the Root Mean Square Error (RMSE), baseline_4 again exhibits the best performance with a value of 0.912046, suggesting minimal overall prediction errors. Baseline_3 maintains strong performance here as well, with an RMSE of 1.076227.
Examining the Hit Rate, baseline_3 leads the pack with 9.35%, signifying a higher success rate in recommendations compared to the other baselines. Meanwhile, baseline_1 and baseline_4 show lower hit rates at 5.61% and 7.48% respectively.
Lastly, looking at the Novelty metric, baseline_4 scores the lowest at 60.35, indicating that its recommendations are less novel or more conventional compared to the others. On the other hand, baseline_1 scores the highest in novelty at 99.41, implying that its recommendations are more diverse or less conventional.
In summary, baseline_4 appears to excel in several metrics including MAE, RMSE, and maintaining relatively low novelty. Baseline_3 stands out with a higher hit rate, showcasing effectiveness in recommendation success. Baseline_2, despite not excelling in the other metrics, exhibits an exceptionally high novelty score, indicating a unique approach to recommendations compared to the rest.
......
0% ou .
You are about to add 0 people to the discussion. Proceed with caution.
Terminez d'abord l'édition de ce message.
Veuillez vous inscrire ou vous pour commenter