This repo publish the code for the ESWC 2020 paper SAShA: Semantic-Aware Shilling Attacks on Recommender Systems Exploiting Knowledge Graphs.

Table of Contents:

• Requirements
• Datasets
• Reproducibility Details
  • 1. Identify the Target Items
  • 2. Perform the Shilling Attacks
  • 3. Evaluation

To run the experiments, it is necessary to install the following requirements.

• Python 3.6.9
• CUDA 10.1
• cuDNN 7.6.4

After having clone this repository with

git clone repo-name

we suggest creating e virtual environment install the required Python dependencies with the following commands

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

The tested datasets across the paper experiments are reported in the following table.

The training dataset is available at ./data/<dataset_name>/ratings.csv with the format

user-Id,tem-Id,rating

The file ./data/<dataset_name>/df_map.csv contains the connection between items and features in the format

feature,item,item_index,value

while, the file ./data/<dataset_name>/features.tsv contains the predicate and object URIs related to each indexed feature, e.g., 0[TAB]<http://dbpedia.org/ontology/publisher><http://dbpedia.org/resource/Pocket_Books>,

Additionally, the file ./data/<dataset_name>/selected_features.csv contains the features filtered with the process define din the article. It is possible to put here the list of features saved following the next template:

features,type
"[1619, 2133, 5092, 10048, 39949, 3235, 94, 33182]",categorical
"[1, 5, 98, 7, 465, 1025]",factual
"[0, 8999, 4672]",ontological

The datasets used in the experiments ara available in the current repo by downloading and extracting this archive.

To execute the following script it is necessary to execute the following command in the shell

cd application

The first step is to generate the target items executing the following command

python run_generate_target_items.py 
  --datasets <data-name-1> <data-name-2> 
  --num_target_items <num_target_items> 

The script will generate the file ./data/<dataset_name>/target_items.csv. We have performed our experiments setting num_target_items = 50.

After having executed the previous commands we can start the flow of attacking the recommender models by running the following command.

python run_multiple_execution_server.py 
         --gpu  -1
         --random_sampling      1
         --initial_predictions  0
         --evaluate_similarities        0
         --generate_profiles    0
         --post_predictions     0
         --similarity_types     ['katz', 'exclusivity', 'cosine']
         --semantic_attack_types        ['target_similar']
         --topk 10
         --alpha        0.25
         --models       ['NCF', 'SVD', 'ItemkNN', 'UserkNN']
         --selection_types      ['categorical', 'ontological', 'factual']
         --datasets     ['SmallLibraryThing']
         --item_size    0.05
         --number_processes     1
         --attacks      ['Random', 'Average']
         --num_target_items     10
         --top_k_similar_items  0.25
         --size_of_attacks      [0.01, 0.025, 0.05]
         --station      <server-name>

Note that we have provided an example of execution. It is possible to change the command parameters to reproduce/execute any type of attack. At the end of the execution all the crafted shilling profiles will be store in ./<model-name>/shilling_profiles/<dataset_name>/ while the positions and scores of the Target Items will be saved in ./<model-name>/results/<dataset_name>/.

To measure all the results that we report in tha paper, we can execute the following command

python run_process_attack_result.py 
         --datasets     ['SmallLibraryThing']
         --models       ['NCF']
         --metrics      ['HR', 'PS']
         --top_k_metrics        10
         --semantic_attack_types        ['target_similar', 'baseline']
         --attacks      ['Random', 'Average']
         --station      not-specified

At the end of this command, the results files will be saved in ./<model-name>/HR/<dataset_name>/ and ./<model-name>/PS/<dataset_name>/ directories. Note that, the results are evaluated within a significance statistical test by adding asterisks at the end of each PS/HR values.

@inproceedings{DBLP:conf/esws/AnelliDNSM20,
  author    = {Vito Walter Anelli and
               Yashar Deldjoo and
               Tommaso Di Noia and
               Eugenio Di Sciascio and
               Felice Antonio Merra},
  title     = {SAShA: Semantic-Aware Shilling Attacks on Recommender Systems Exploiting
               Knowledge Graphs},
  booktitle = {{ESWC}},
  series    = {Lecture Notes in Computer Science},
  volume    = {12123},
  pages     = {307--323},
  publisher = {Springer},
  year      = {2020}
}