136 6. MULTIMODAL SEQUENTIAL LEARNING
In addition, we also conducted the significance test between our model and the most com-
petitive baseline THACIL. We can see that the advantage of our model is statistically significant
as p-value is 2:81 10
5
on the Dataset III-1 and 4:70 10
6
on the Dataset III-2.
To justify the robustness of our proposed model, we comparatively explored the perfor-
mance of our model and the baselines by varying the number of returned items K. Figure 6.5
shows the results regarding the performance comparison on K.
• Jointly analyzing the performance of the models in Figures 6.5a and d, we found that
increasing the number of returned items K degrades the precision value of the recommen-
dation. But our model ALPINE outperforms others under the same experimental setting,
especially on the Dataset III-1.
• e performance of all these methods over recall and F value rises fast as the number of
returned items K linearly increases. eir curves then gradually ascend to a steady state.
Our method ALPINE consistently and remarkably outputs a higher accuracy as compared
to that of other methods, especially on the Dataset-III-2. is verifies the robustness of
our model.
6.5.4 COMPONENT-WISE EVALUATION OF ALPINE
We studied the variants of our model to further investigate the effectiveness of the uninterested
representation modeling, user-matrix, and temporal interest graph:
• ALPINE_u: We eliminated the uninterested representation modeling part from the
model. Namely, we computed the final click probability by interested representation and
multi-level interest representation.
• ALPINE_m: We eliminated the multi-level interest module. at is, the final click prob-
ability is computed by the user’s interested and uninterested representation.
• ALPINE_um: We only utilized the user’s interested sequence to predict the click proba-
bility, namely we eliminated both the uninterested representation modeling and the multi-
level interest modeling layer.
• ALPINE_umg: We eliminated the graph information from the ALPINE_um model.
We compared these variants on the two datasets, and Table 6.2 summarizes the results
regarding the component-wise comparison. By jointly analyzing Table 6.2, we gained the fol-
lowing insights:
• By jointly analyzing the performance of ALPINE_u on the two datasets, it can be seen that
removing the uninterested representation modeling degrades the recommendation results.
To be more specific, ALPINE_u has dropped by 0.2% on the Dataset III-1 and 1.1%
on the Dataset III-2 in terms of AUC. is verifies the effectiveness of the uninterested
representation modeling.