replicating the results

[majidhosseini87]

Hi,
Thank you for your amazing efforts. I've been trying to replicate the results of the paper titled "PSAC: Proactive Sequence-Aware Content Caching via Deep Learning at the Network Edge" using your code. Unfortunately, I am facing challenges in achieving the results described in the paper. Specifically, the results from the Psac_gen framework in your code significantly differ from the Qoe_score mentioned paper. Could you provide any guidance or updates that might assist in accurately replicating the results?

Sincerely,

[DarriusL]

Hi,
PSAC_gen in CoCheLab is mainly used for comparison with contrastive learning models (CL4SRec, etc.), so PSAC_gen here is a modified version:

1. The user request sequence in the original paper was not clipped or filled to a fixed length. In order to unify it with other models, PSAC_gen here is also trained with a fixed-length sequence.
2. The QoE in the original paper is calculated as follows:

, where \theta is the average length of the sequence. Due to the former reason, if the original formula is used, the QoE must be 0 (fixed-length sequence). Therefore, I made some modifications to the original formula: , where \theta is the user satisfaction rate and is a hyperparameter.

[Farzad-Mehrabi]

hey, in PSAC Framework for the forward pass there is tensor called su, could you define exactly what it is and what slide_len and L represent?
def forward(self, su):
#su: (batch_size, slide_len, L)
#Ec: (batch_size, slide_len, L, d)
Ec = self.encoder(su);
#Eu: (batch_size, 1, d)
Eu = self.encoder(su.reshape(su.shape[0], -1)).mean(dim = 1).unsqueeze(1);
#o: (batch_size, slide_len, 1, n)
o = self.VrtConv(Ec.transpose(0,1)).transpose(0,1).transpose(-1, -2);
#attn: (batch_size, slide_len, 1, L*d)
attn = self.self_attn(Ec);
#pro_logits: (batch_size, slide_len, req_set_len)
pro_logits = self.LSTFcNet(o, attn, Eu);
return pro_logits

[DarriusL]

hey, in PSAC Framework for the forward pass there is tensor called su, could you define exactly what it is and what slide_len and L represent? def forward(self, su): #su: (batch_size, slide_len, L) #Ec: (batch_size, slide_len, L, d) Ec = self.encoder(su); #Eu: (batch_size, 1, d) Eu = self.encoder(su.reshape(su.shape[0], -1)).mean(dim = 1).unsqueeze(1); #o: (batch_size, slide_len, 1, n) o = self.VrtConv(Ec.transpose(0,1)).transpose(0,1).transpose(-1, -2); #attn: (batch_size, slide_len, 1, L*d) attn = self.self_attn(Ec); #pro_logits: (batch_size, slide_len, req_set_len) pro_logits = self.LSTFcNet(o, attn, Eu); return pro_logits

sure. I usually use su to represent the user sequence, and the input in PSAC_gen ([batch, slide_len, L]) is the original user sequence ([batch, n]) processed by the sliding window (length L)