Page 173 - The-5th-MCAIT2021-eProceeding

creates user-item embeddings and computes user-item similarity matching scores (X. He, 2017). The
representation learning approach learns much richer user-item representations (Wu. L, 2019). The advantages
of using DL for representation learning are twofold: (1) it reduces the efforts in handcraft feature design; and
(2) enables recommendation models to include heterogeneous content information such as text, images, audio,
and video. Deep representation learning helps to improve the search and recommendation performance but lacks
transparency and explainability. The recommendations are hardly explainable to system designers/users.

2.2.3. Knowledge Graph
After LFM, the next big effort in the recommendation engine is the applicability of the KG. The KG improves
prediction accuracy and enhances the explainability of the models. KG embeddings (KGE) have started leading
the effort in explainable recommendation engines. Few researchers start focusing on TransE (Bordes, 2013)
and node2vec (Grover, 2016) models. The objective of KGE is to find the similarity between entities by
calculating their representation distance (Zhang F., 2016). The main issue with KGE models is that they produce
black-box recommendations. Ai et al. 2018 enhanced the CF approach over KGE and later adopted a soft
matching algorithm to find explanation paths between users and items. Here the explanation is not produced
according to the reasoning process, but it is a post-hoc explanation. It is computationally expensive to traverse
all the paths between a user-item node pair for similarity calculation in real world KG.

2.2.4. Reinforcement Learning
RL is another big area, and many successful applications belong to this area e.g. AlphaGo (D Silver, 2016),
self-driving cars (Chopra, 2020) etc. It works under the framework of State, Agent, and Actions. In every state,
the agent takes an appropriate action to maximize the rewards and go to another state. This framework belongs
to an environment and demonstrates the ability of the agent to understand the high-level causal relationships.
Xian et al., 2019 proposed a RL model for pathfinding to address the KG traversal issue for explainable
recommendations. In the training stage, the agent finds the right user-item path with high rewards. In the
inference stage, the agent will use the high reward path for recommendations. It thus addresses the issue of
enumerating all the paths between user-item pairs. Using RL for recommendation is a fairly novel field.

2.2.5. Language Models
Large-scale pre-trained language models like OpenAI GPT and Bidirectional Encoder Representations from
Transformer (BERT) have achieved great performance on a variety of language tasks using generic model
architectures. Language models like BERT, GPT, ELMO, and many others are the main drivers for having
semantic implementations and have a great scope in the recommendation engine (Shaoxiong, 2021). (Fei Sun,
2019) explained the sequential recommendation by using BERT. The research of generating natural language
(NLG) explanation is still in its early stage, and LM have a great scope in enhancing the same.

3. Conclusion

Explainable recommendation, which provides explanations about the recommendations made to the user,
has attracted increasing attention due to its ability in helping users make better decisions and increasing users’
trust in the system. The idea of this paper is to understand various implementation approaches, their merits, and
demerits for the successful implementation of an explainable recommendation engine.
Content and CF algorithms are the early approaches for developing recommendation engines. CB algorithms
define user preferences based on the content information and the user’s past interaction with the system. The
model can capture the user-specific interests and recommend niche items. The main demerit with this algorithm
is the domain-specific nature of the recommendation. CF algorithms have addressed this domain-specific issue.
This algorithm is mainly using the wisdom of the crowd. The main demerits associated with CF algorithms are
the cold-start problem and lesser explainability power than CB algorithms.

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [160]
Artificial Intelligence in the 4th Industrial Revolution