PO4ISR

PO4ISR is a simple yet effective paradigm for Intent-aware session recommendation (ISR) motivated by the advanced reasoning capability of large language models (LLMs). Specifically, we first create an initial prompt to instruct LLMs to predict the next interacted item by inferring varying user intents reflected in a session. Then, an effective prompt optimization mechanism is proposed to automatically optimize prompts with iterative self-reflection. Finally, the prompt selection module is designed to efficiently choose optimized prompts, leveraging the robust generalizability of LLMs across diverse domains.

Installtion

Install all dependencies via:

pip install -r requirements.txt

Datasets

We adopt three real-world datasets from various domains: MovieLens-1M, Games and Bundle. You can find the datasets used in the experiment under the Dataset directory. Each dataset includes both its ID and text format. In addition to providing the randomly-sampled training data, we also provide full versions of the training data under the Dataset directory.

ID-Formatted Dataset

• train_sample_x.npy: randomly select x sessions from the full version of the training dataset as the training set. x can be 50 or 150.
• train.npy: full version of the training data.
• valid.npy: the validation set containing all validation sessions.
• valid_candidate.npy: the candidate set corresponding to each session in the validation set.
• test.npy: test set containing all test sessions.
• test_candidate_x.npy: the candidate sets constructed by 5 different random seeds, corresponding to each session in the test set. x can be 0, 10, 42, 625 or 2023.

Text-Formatted Dataset

• train_x.json: the training set in text format corresponding to the train_sample_x.npy file in ID format.
• valid.json: the validation set in text format containing both validation sessions and the candidates, which corresponds to the valid.npy and valid_candidate.npy file in ID format.
• test_seed_x.json: the test set in text format containing both test sessions and the candidates, which corresponds to the test.npy and test_candidate_x.npy file in ID format.

PO4ISR

Tune

The tune.py corresponds to the process of prompt optimization. Before running the code, you need to fill in your OpenAI API token in the ./PO4ISR/assets/openai.yaml file and wandb token in the ./PO4ISR/assets/overall.yaml file.

python tune.py --dataset='dataset name' --sample_num='number of training data'

Test

The test.py file corresponds to the evaluation of the prompt.

python test.py --dataset='dataset name' --seed='value of the seed'

Note that all the optimal prompts are saved in the PO4ISR/prompts.py file. If you want to test the results with these prompts, you can replace them in the test.py.

NIR

python test.py --dataset='dataset name' --seed='value of the seed' --api_key='your OpenAI API token'

Non-LLM-Baselines

Parameter Tuning and Settings for Non-LLM-Baselines

We use the open-source framework Optuna to automatically find out the optimal hyperparameters of all methods with 50 trails. The item embedding size is searched from {32, 64, 128}; learning rate is searched from {10−4, 10−3, 10−2}; batch size is searched from {64, 128, 256} and we use an early stop mechanism to halt the model training, with a maximum of 100 epochs. For SKNN, 𝐾 is searched from {50, 100, 150}. For NARM, the hidden size is searched in [50, 200] stepped by 50, and the number of layers is searched in {1, 2, 3}. For GCE-GNN, the number of hops is searched in {1, 2}; the dropout rate for global aggregators is searched in [0, 0.8] stepped by 0.2 and the dropout rate for local aggregators is searched in {0, 0.5}. For MCPRN, 𝜏 is searched in {0.01, 0.1, 1, 10} and the number of purpose channels is searched in {1, 2, 3, 4}. For HIDE, the number of factors is searched in {1, 3, 5, 7, 9}; the regularization and balance weights are searched in {10−5, 10−4, 10−3, 10−2}; the window size is searched in [1, 10] stepped by 1; and the sparsity coefficient is set as 0.4. For Atten-Mixer, the intent level 𝐿 is searched in [1, 10] stepped by 1 and the number of attention heads is searched in {1, 2, 4, 8}. The optimal parameter settings are shown in Table 1.

                  Table 1: Optimal Parameter Settings for Non-LLM-Baselines.

	Bundle	ML-1M	Games
SKNN
FPMC
NARM
STAMP
GCE-GNN
MCPRN
HIDE
Atten-Mixer

Tune

You can run the following command to tune the model and find the optimal parameter combination.

python tune.py --dataset='dataset name' --sample_num='number of training data' --model='model name'

Test

After completing the tuning process, we can find the optimal parameter settings in the tune_log directory. Additionally, we have placed the optimal parameters obtained during the experiment in the config.py. You can also use the following command to directly test the model.

python test.py --dataset='dataset name' --model='model name' --seed='value of the seed'

Acknowledgment

We refer to the following repositories to implement the baselines in our code:

• Non-LLM-Baselines part with Understanding-Diversity-in-SBRSs
• LLM-Baseline-NIR parts with LLM-Next-Item-Rec