introduction_to_recommender_systems/P5. Graph-based.ipynb

Self made RP3-beta

import helpers
import pandas as pd
import numpy as np
import scipy.sparse as sparse
from collections import defaultdict
from itertools import chain
import random
import time
import matplotlib.pyplot as plt

train_read = pd.read_csv("./Datasets/ml-100k/train.csv", sep="\t", header=None)
test_read = pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None)
(
    train_ui,
    test_ui,
    user_code_id,
    user_id_code,
    item_code_id,
    item_id_code,
) = helpers.data_to_csr(train_read, test_read)

class RP3Beta:
    def fit(self, train_ui, alpha, beta):
        """We weight our edges by user's explicit ratings so if user rated movie high we'll follow that path
        with higher probability."""
        self.train_ui = train_ui
        self.train_iu = train_ui.transpose()

        self.alpha = alpha
        self.beta = beta

        # Define Pui
        Pui = sparse.csr_matrix(self.train_ui / self.train_ui.sum(axis=1))

        # Define Piu
        to_divide = np.vectorize(lambda x: x if x > 0 else 1)(
            self.train_iu.sum(axis=1)
        )  # to avoid dividing by zero
        Piu = sparse.csr_matrix(self.train_iu / to_divide)
        item_orders = (self.train_ui > 0).sum(axis=0)

        Pui = Pui.power(self.alpha)
        Piu = Piu.power(self.alpha)

        P3 = Pui * Piu * Pui

        P3 /= np.power(
            np.vectorize(lambda x: x if x > 0 else 1)(item_orders), self.beta
        )

        self.estimations = np.array(P3)

    def recommend(self, user_code_id, item_code_id, topK=10):

        top_k = defaultdict(list)
        for nb_user, user in enumerate(self.estimations):

            user_rated = self.train_ui.indices[
                self.train_ui.indptr[nb_user] : self.train_ui.indptr[nb_user + 1]
            ]
            for item, score in enumerate(user):
                if item not in user_rated and not np.isnan(score):
                    top_k[user_code_id[nb_user]].append((item_code_id[item], score))
        result = []
        # Let's choose k best items in the format: (user, item1, score1, item2, score2, ...)
        for uid, item_scores in top_k.items():
            item_scores.sort(key=lambda x: x[1], reverse=True)
            result.append([uid] + list(chain(*item_scores[:topK])))
        return result

    def estimate(self, user_code_id, item_code_id, test_ui):
        result = []
        for user, item in zip(*test_ui.nonzero()):
            result.append(
                [
                    user_code_id[user],
                    item_code_id[item],
                    self.estimations[user, item]
                    if not np.isnan(self.estimations[user, item])
                    else 1,
                ]
            )
        return result

model = RP3Beta()
model.fit(train_ui, alpha=1, beta=0)

top_n = pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))

top_n.to_csv(
    "Recommendations generated/ml-100k/Self_P3_reco.csv", index=False, header=False
)

estimations = pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
estimations.to_csv(
    "Recommendations generated/ml-100k/Self_P3_estimations.csv",
    index=False,
    header=False,
)

import evaluation_measures as ev

estimations_df = pd.read_csv(
    "Recommendations generated/ml-100k/Self_P3_estimations.csv", header=None
)
reco = np.loadtxt("Recommendations generated/ml-100k/Self_P3_reco.csv", delimiter=",")

ev.evaluate(
    test=pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None),
    estimations_df=estimations_df,
    reco=reco,
    super_reactions=[4, 5],
)

943it [00:00, 8787.46it/s]

	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
0	3.702446	3.527273	0.282185	0.192092	0.186749	0.21698	0.204185	0.240096	0.339114	0.204905	0.572157	0.593544	0.875928	1.0	0.077201	3.875892	0.974947

Let's check hyperparameters

Alpha

from tqdm import tqdm

result = []
for alpha in tqdm([round(i, 1) for i in np.arange(0.2, 1.6001, 0.2)]):
    model = RP3Beta()
    model.fit(train_ui, alpha=alpha, beta=0)
    reco = pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))
    estimations_df = pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
    to_append = ev.evaluate(
        test=pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None),
        estimations_df=estimations_df,
        reco=np.array(reco),
        super_reactions=[4, 5],
    )
    to_append.insert(0, "Alpha", alpha)
    result.append(to_append)

result = pd.concat(result)
result

  0%|          | 0/8 [00:00<?, ?it/s]
943it [00:00, 9826.90it/s]
 12%|█▎        | 1/8 [00:07<00:55,  7.93s/it]
0it [00:00, ?it/s]
943it [00:00, 8050.54it/s]
 25%|██▌       | 2/8 [00:15<00:47,  7.94s/it]
943it [00:00, 9441.40it/s]
 38%|███▊      | 3/8 [00:24<00:40,  8.01s/it]
943it [00:00, 9745.57it/s]
 50%|█████     | 4/8 [00:32<00:32,  8.13s/it]
943it [00:00, 10134.88it/s]
 62%|██████▎   | 5/8 [00:40<00:23,  7.98s/it]
943it [00:00, 10252.12it/s]
 75%|███████▌  | 6/8 [00:47<00:15,  7.94s/it]
943it [00:00, 9769.78it/s]
 88%|████████▊ | 7/8 [00:55<00:07,  7.97s/it]
0it [00:00, ?it/s]
943it [00:00, 8272.37it/s]
100%|██████████| 8/8 [01:03<00:00,  7.99s/it]

	Alpha	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
0	0.2	268.177832	211.732649	0.262672	0.166858	0.166277	0.197184	0.187661	0.203252	0.320910	0.196132	0.563378	0.580866	0.850477	1.000000	0.060606	3.669627	0.979636
0	0.4	10.546689	7.792373	0.268505	0.172669	0.171569	0.202643	0.192489	0.212653	0.326760	0.200172	0.565148	0.583801	0.854719	1.000000	0.064214	3.726996	0.978426
0	0.6	3.143988	2.948790	0.274655	0.180502	0.177820	0.208730	0.198176	0.222746	0.332872	0.203290	0.568872	0.587738	0.870626	1.000000	0.065657	3.785282	0.977090
0	0.8	3.670728	3.495735	0.281972	0.189868	0.185300	0.216071	0.203541	0.236751	0.339867	0.206688	0.573729	0.592432	0.874867	1.000000	0.070707	3.832415	0.975998
0	1.0	3.702446	3.527273	0.282185	0.192092	0.186749	0.216980	0.204185	0.240096	0.339114	0.204905	0.572157	0.593544	0.875928	1.000000	0.077201	3.875892	0.974947
0	1.2	3.704441	3.529251	0.280912	0.193633	0.187311	0.216872	0.203004	0.240588	0.338049	0.203453	0.571830	0.594313	0.883351	1.000000	0.085859	3.910718	0.974073
0	1.4	3.704580	3.529388	0.273595	0.190651	0.183874	0.212183	0.199464	0.239118	0.329550	0.195433	0.566171	0.592793	0.871686	1.000000	0.107504	3.961915	0.972674
0	1.6	3.704591	3.529399	0.263097	0.186255	0.178709	0.205170	0.191094	0.232920	0.317439	0.184917	0.552349	0.590545	0.868505	0.999576	0.156566	4.060156	0.969203

metrics = list(result.columns[[i not in ["Alpha"] for i in result.columns]])

charts_per_row = 6
charts_per_column = 3

fig, axes = plt.subplots(
    nrows=charts_per_row, ncols=charts_per_column, figsize=(18, 7 * charts_per_row)
)
import itertools

to_iter = [
    i for i in itertools.product(range(charts_per_row), range(charts_per_column))
]

for i in range(len(metrics)):
    df = result[["Alpha", metrics[i]]]
    df.plot(ax=axes[to_iter[i]], title=metrics[i], x=0, y=1)

Beta

from tqdm import tqdm

result = []
for beta in tqdm([round(i, 1) for i in np.arange(0, 1, 0.1)]):
    model = RP3Beta()
    model.fit(train_ui, alpha=1, beta=beta)
    reco = pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))
    estimations_df = pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
    to_append = ev.evaluate(
        test=pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None),
        estimations_df=estimations_df,
        reco=np.array(reco),
        super_reactions=[4, 5],
    )
    to_append.insert(0, "Beta", beta)
    result.append(to_append)

result = pd.concat(result)
result

  0%|          | 0/10 [00:00<?, ?it/s]
0it [00:00, ?it/s]
943it [00:00, 8133.31it/s]
 10%|█         | 1/10 [00:08<01:19,  8.87s/it]
0it [00:00, ?it/s]
943it [00:00, 8733.83it/s]
 20%|██        | 2/10 [00:17<01:10,  8.81s/it]
943it [00:00, 9872.77it/s]
 30%|███       | 3/10 [00:25<01:00,  8.61s/it]
943it [00:00, 10989.65it/s]
 40%|████      | 4/10 [00:33<00:50,  8.36s/it]
943it [00:00, 10108.59it/s]
 50%|█████     | 5/10 [00:41<00:40,  8.15s/it]
943it [00:00, 9989.31it/s]
 60%|██████    | 6/10 [00:49<00:32,  8.09s/it]
943it [00:00, 10311.97it/s]
 70%|███████   | 7/10 [00:56<00:23,  7.92s/it]
943it [00:00, 10283.20it/s]
 80%|████████  | 8/10 [01:04<00:15,  7.79s/it]
943it [00:00, 10486.98it/s]
 90%|█████████ | 9/10 [01:11<00:07,  7.75s/it]
943it [00:00, 10330.42it/s]
100%|██████████| 10/10 [01:19<00:00,  7.95s/it]

	Beta	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
0	0.0	3.702446	3.527273	0.282185	0.192092	0.186749	0.216980	0.204185	0.240096	0.339114	0.204905	0.572157	0.593544	0.875928	1.000000	0.077201	3.875892	0.974947
0	0.1	3.703312	3.528128	0.290138	0.197597	0.192259	0.223336	0.210944	0.246153	0.347768	0.212034	0.581038	0.596328	0.884411	1.000000	0.085137	3.957416	0.972784
0	0.2	3.703825	3.528636	0.297137	0.201202	0.196067	0.228169	0.218026	0.252767	0.355655	0.219909	0.588904	0.598160	0.886532	1.000000	0.094517	4.053212	0.969980
0	0.3	3.704130	3.528939	0.303499	0.204749	0.199901	0.232829	0.225107	0.260797	0.363757	0.226825	0.599969	0.599964	0.888653	1.000000	0.105339	4.147779	0.966948
0	0.4	3.704313	3.529120	0.308908	0.208811	0.203854	0.237241	0.229614	0.266918	0.370758	0.232673	0.609385	0.602014	0.895016	0.999894	0.132035	4.259682	0.962989
0	0.5	3.704422	3.529229	0.314316	0.211411	0.206768	0.240986	0.237124	0.273416	0.378307	0.239297	0.622792	0.603327	0.903499	0.999046	0.168831	4.411281	0.956648
0	0.6	3.704488	3.529295	0.314634	0.206209	0.204818	0.240159	0.242489	0.273850	0.376438	0.238428	0.622042	0.600721	0.897137	0.996394	0.212843	4.621938	0.945932
0	0.7	3.704528	3.529335	0.304136	0.187298	0.191990	0.228749	0.238305	0.256201	0.358807	0.226808	0.593897	0.591207	0.868505	0.983033	0.256854	4.898568	0.928065
0	0.8	3.704552	3.529360	0.266384	0.147571	0.158660	0.194838	0.214485	0.209336	0.299850	0.184356	0.492852	0.571152	0.803818	0.936373	0.341270	5.257397	0.895882
0	0.9	3.704567	3.529375	0.162354	0.076967	0.089233	0.114583	0.134657	0.113253	0.160868	0.085486	0.243590	0.535405	0.580064	0.800106	0.415584	5.563910	0.857396

metrics = list(result.columns[[i not in ["Beta"] for i in result.columns]])

charts_per_row = 6
charts_per_column = 3

fig, axes = plt.subplots(
    nrows=charts_per_row, ncols=charts_per_column, figsize=(18, 7 * charts_per_row)
)
import itertools

to_iter = [
    i for i in itertools.product(range(charts_per_row), range(charts_per_column))
]

for i in range(len(metrics)):
    df = result[["Beta", metrics[i]]]
    df.plot(ax=axes[to_iter[i]], title=metrics[i], x=0, y=1)

Check sample recommendations

train = pd.read_csv(
    "./Datasets/ml-100k/train.csv",
    sep="\t",
    header=None,
    names=["user", "item", "rating", "timestamp"],
)
items = pd.read_csv("./Datasets/ml-100k/movies.csv")

user = random.choice(list(set(train["user"])))

train_content = pd.merge(train, items, left_on="item", right_on="id")
display(
    train_content[train_content["user"] == user][
        ["user", "rating", "title", "genres"]
    ].sort_values(by="rating", ascending=False)[:15]
)

reco = np.loadtxt("Recommendations generated/ml-100k/Self_P3_reco.csv", delimiter=",")
items = pd.read_csv("./Datasets/ml-100k/movies.csv")

# Let's ignore scores - they are not used in evaluation:
reco_users = reco[:, :1]
reco_items = reco[:, 1::2]
# Let's put them into one array
reco = np.concatenate((reco_users, reco_items), axis=1)

# Let's rebuild it user-item dataframe
recommended = []
for row in reco:
    for rec_nb, entry in enumerate(row[1:]):
        recommended.append((row[0], rec_nb + 1, entry))
recommended = pd.DataFrame(recommended, columns=["user", "rec_nb", "item"])

recommended_content = pd.merge(recommended, items, left_on="item", right_on="id")
recommended_content[recommended_content["user"] == user][
    ["user", "rec_nb", "title", "genres"]
].sort_values(by="rec_nb")

	user	rating	title	genres
39675	599	5	That Thing You Do! (1996)	Comedy
44133	599	5	Father of the Bride Part II (1995)	Comedy
55221	599	5	Fan, The (1996)	Thriller
34957	599	5	My Best Friend's Wedding (1997)	Comedy, Romance
34654	599	5	Sabrina (1995)	Comedy, Romance
56202	599	5	Happy Gilmore (1996)	Comedy
57099	599	5	Up Close and Personal (1996)	Drama, Romance
60181	599	5	One Night Stand (1997)	Drama
5385	599	5	Jerry Maguire (1996)	Drama, Romance
12546	599	5	Truth About Cats & Dogs, The (1996)	Comedy, Romance
9903	599	5	Time to Kill, A (1996)	Drama
8705	599	5	To Gillian on Her 37th Birthday (1996)	Drama, Romance
63190	599	4	Craft, The (1996)	Drama, Horror
60634	599	4	Romy and Michele's High School Reunion (1997)	Comedy
63675	599	4	Set It Off (1996)	Action, Crime

	user	rec_nb	title	genres
262	599.0	1	Star Wars (1977)	Action, Adventure, Romance, Sci-Fi, War
7163	599.0	2	Fargo (1996)	Crime, Drama, Thriller
2874	599.0	3	Return of the Jedi (1983)	Action, Adventure, Romance, Sci-Fi, War
4627	599.0	4	Air Force One (1997)	Action, Thriller
5873	599.0	5	Contact (1997)	Drama, Sci-Fi
8551	599.0	6	Independence Day (ID4) (1996)	Action, Sci-Fi, War
1189	599.0	7	English Patient, The (1996)	Drama, Romance, War
8274	599.0	8	Mr. Holland's Opus (1995)	Drama
7633	599.0	9	Rock, The (1996)	Action, Adventure, Thriller
2464	599.0	10	Godfather, The (1972)	Action, Crime, Drama

project task 5: generate recommendations of RP3Beta for hyperparameters found to optimize recall

# We generated recommendations for P3, a special case of RP3Beta (with alpha=1, beta=0).
# We've observed that changing alpha and beta impacts the model performance.

# Your task is find values alpha and beta for which recall will be the highest 
# (any solution with recall higher than P3 will be accepted)
# train the model and generate recommendations.

# save the outptut in 'Recommendations generated/ml-100k/Self_RP3Beta_estimations.csv'
# and 'Recommendations generated/ml-100k/Self_RP3Beta_reco.csv'

project task 6 (optional): implement graph-based model of your choice

# for example change length of paths in RP3beta or make some other modification (but change more than input and hyperparameters)
# feel free to implement your idea or search for some ideas

# save the outptut in 'Recommendations generated/ml-100k/Self_GraphTask_estimations.csv'
# and 'Recommendations generated/ml-100k/Self_GraphTask_reco.csv'