Rating each user and item by iterating a behavior matrix.

• user score: $\mathbf e = (e_1, \cdots, e_u, \cdots, e_M)^\top \in \mathbb R^M $
• item score: $\mathbf q = (q_1, \cdots, q_u, \cdots, q_N)^\top \in \mathbb R^N $
• behavior matrix: $A = {A_{ui}} \in \mathbb R^{M\times N}$
• in iteration $t$: $$ \mathbf e^{(t)} = \frac 1 Z A \ \mathbf q^{(t-1)} \ \mathbf q^{(t)} = \frac 1 Z A^\top \ \mathbf e^{(t)} $$ where $Z$ is for normalization. In this implementation, $Z$ takes the max value (inducing max score tobe 1).

configure in file ./src/config.ini

• iteration: max num of iteration for train
• stop_err: threthold value of RSE difference of two adjacent iterations to stop train
• top_num: number of top users(Experts) and top items(high quality items) to output

• training data file is setted in config file: file: train_data
• format (each line): uid pid score

• user score: file: user_score
• item score: file: item_score
• top users(Expert): file: top_user
• top items: file: top_item

you can use ./src/demo.sh to for a complete demonstration

cd ./lib; make clean; make
cd ./src; make clean; make

cd ./src
./generate_data.py

cd src
./spear config.ini

SPEAR: SPAMMING-RESISTANT EXPERTISE ANALYSIS AND RANKING IN COLLABORATIVE TAGGING SYSTEMS