SMPCUP2017

Team: ELP

Paper link: http://www.data-intelligence.org/p/19/

Organization：University of International Relations

Members：Lu Junru; Chen Le; Meng Kongming; Wang Fengyi; Xiangjun; Zhou Kaimin; Dong Zhenyuan; Shan jiawei; Lian Lingchen

Introduction:

this repository is established to review codes and documents for our team ELP in SMPCUP2017, which is a user profiling contest based on massive data provided by CSDN, including text, relations and interactions of users.

In SMPCUP2017, every team is requested to work on three specific tasks:

TASK1: Given a number of user documents (blogs or posts), generate 3 most appropriate keywords for each document. The generated keywords must appear in the document.
TASK2: Given the user's document information (blog or post) and behavior data (browsing, comment, collection, forwarding, point-of-thumb, step-by-step, private, etc.) for each user, mark the three most relevant interests for each user. The label space is given by CSDN.
TASK3: Given a number of users in a period of time (at least 1 year) document information (blog or post) and behavioral data (browsing, commentary, collection, forwarding, dating, attention, private messages, etc.), predict each user in the future over a period of time (half a year to 1 year) growth value. User growth is based on the user's overall performance scoring income, but will not publish the specific score criteria. The growth value will be normalized to the [0, 1] interval, where the value is 0 for user churn.

More detailed imformation could be browsed on the page: https://biendata.com/competition/smpcup2017/

Baseline Models:

TASK1: TFIDF(Reference: https://en.wikipedia.org/wiki/Tf–idf#Term_frequency)
TASK2: W-BAG, a model based on Word2Vec and BaggingClassifer
TASK3: BPNN-XGboost(BPXG), a linear average model of BackPropagationNeuralNetwork and XGboost

Final Models:

TASK1: S-TFIDF, a promoted model based on TFIDF and Textrank.
TASK2: S-TFIDF/DocumentEmbedding-SVC-Stacking(SDSS), a stacking model that using S-TFIDF and DocumentEmbedding as first layer and using SVC as second layer.
TASK3: PAR/GDR-NuSVR-Stacking(PGNS), a stacking model that using PassiveActiveRegressor and GrandientBoostingRegressor as first layer and using NuSVR as second layer.

Performance:

TASK1	TRAIN	VALID	TEST
TFIDF	0.56	0.52	None
S-TFIDF	0.61	0.56	0.56

TASK2	TRAIN	VALID	TEST
W-BAG	None	0.40	0.373
SDSS	None	0.39	0.378

TASK3	TRAIN	VALID	TEST
BPXG	0.54	0.59	None
PGNS	0.765	0.73	0.75

Environment:

Task1: python 2.7
Task2: python 3.0
Task3: python 2.7

LuJunru/SMPCUP2017_ELP